• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    Method for classifying a two- or higher-dimensional image, which comprises elements N-dimensional data set is made to represent the image in that at least two of the N dimensions constitute respective axes in the image, so that a certain pixel in the image corresponds to a certain element in the data set , and in that each element is associated with M numerical care, each of which represents a measure of a property of the element in question, of which at least one property represents image information in the respective sampled channel. The invention is characterized by the method identifying, firstly, a certain predetermined, variable geometric structure, the extent of which in at least two of the dimensions of the data set is determined in relation to an individual element in the data set and of at least one variable parameter, and secondly at least one geometric dimension associated with said variable geometric structure, which geometric dimensions are arranged to feed a structure in the geometric property of a specific geometric relationship to other specific such geometric structures, and in that the method further comprises step night a computer or several interconnected computers a) digitally mounted on a digit storage medium store the amount of data (10l); b) for at least each element corresponding to a pixel in the image, (N31 for at least one of the P4 element properties, firstly is determined to determine a specific geometric structure which can be obtained in relation to the element in question and by selecting parameter value or parameter value, for which specific geometric structure at least one of said geometric measures is a maximum at the same time as it is geometrically bounded by individual elements or groups of elements whose element property meets a certain predetermined criterion, and secondly is made to calculate at least one of said geometric c) in a database is stored an association between each element and the element has at least one calculated geometric dimension for each property for which a specific geometric structure has been calculated in step b) (103); d) is made to perform a main classification of the data set, which main classification is made to be based on a comparison measure between two elements' respective sets of associated geometric measures, each element being made to be associated with a certain class in a main set of classes (104); and e) stored on a digital storage medium1. the image in step d) is classified into a set of pixels with the same dimensions as the image, each pixel being associated with the class in the main set of classes with which element corresponding to the pixel in the image is associated in step d) (105); and in that said variable geometric structure, dimension and criterion is not selected so that in step b) the calculation measure constitutes only a measure of the smallest distance between the element in question and the nearest element associated with the element property in question. Application text document 2010-09-10 100172EN
    公开号:SE1050937A1
    申请号:SE1050937
    申请日:2010-09-10
    公开日:2012-03-11
    发明作者:Anders Brun；Zihan Hans Liu；Anders Waestfelt；Bo Malmberg；Michael Nielsen
    申请人:Choros Cognition Ab；
    IPC主号:
    专利说明:

    15 20 25 30 three-dimensional MRI of a person; and identification of material deviations based on a photograph of a manufactured detail.
    For time reasons, it is often desirable to provide an automatic interpretation of an image. One way of performing such an interpretation is based on a digital image in two or more dimensions, made up of a number of individual pixels. Each pixel is then associated with a particular pixel class, selected from a number of such classes, the purpose of which is to represent a particular type of pixel. When all pixels have been associated with. a respective class, each class can be inventoried to in this way get an overall picture of where in the picture pixels of a certain type occur. Such a procedure is referred to herein as "classification". A certain class can thus, for example, represent "path", "muscle tissue" or "material defect".
    Typically, classification techniques are used to locate objects and boundaries, in an image. such as lines, curves, fields, etc. Several attempts have been made to provide a method for automatically performing useful classifications of images where knowledge of the image content is limited prior to the classification.
    For example, a method has been proposed in which a moving "window" is caused to move over the image and thus try to characterize a pixel located in the middle of the window to a certain class affiliation by using statistical methods to study the surrounding pixels of the ndttpixel (kernel-based segmentation). For some types of images, such a procedure may be effective, but the classification often becomes very fragmented with classes that include pixels from many different parts of the image. This means that it becomes difficult to interpret useful information from the classification without much manual work.
    An automatic classification of an image that also takes into account all the pixels in an iterative procedure (window-independent classification) has also been proposed. An example of an algorithm that can be used in such a method is a K-mean type cluster analysis. Even such procedures often result in fragmented classifications when used to classify digitally stored images.
    In the article “Automated Segmentation of MR Images of Brain Tumors”, Kaus, Michael R., et al., Radiology 2001; 2l8: 586-591, describes an iterative classification of a three-dimensional MRI image of a human skull.
    The classification is performed iteratively, using local segmentation strategies and a distance transform that calculates the distance between a certain voxel (a three-dimensional pixel) and a certain class, and on the basis of information regarding the grayscale intensity of the voxels included in the image.
    Such a procedure largely consists of steps that must be performed manually in order to achieve sufficient reliability in the finished classified result.
    In addition, a relatively large knowledge of the object is required before the classification begins, for example a comparison image that illustrates a "normal case" or equivalent.
    The Swedish patent application with number 0950140-4 describes a method according to which the pixels in an image first undergo a first classification, after which the smallest geometric application text 2010-0940 100172EN 10 15 20 25 30 the distance between each pixel and a pixel belonging to a certain class is calculated . A second classification 'is then performed * on the basis of these calculated minimum distances.
    The present invention solves the problems described above.
    Thus, the invention relates to a method for classifying a two- or high-dimensional image, in which an N-dimensional data set comprising elements is made to represent the image by at least two of the N dimensions constituting respective axes in the image, so that a certain pixel in the image corresponds to a certain element in the data set, and in that each element is associated with M numerical values, each of which represents a measure of a property of the element in question, of which at least one property represents image information in a respective sampled channel, characterized in that the method comprises to identify certain predetermined, variable geometric firstly a structure, the extent of which in at least two of the N dimensions of the data set is determined in relation to an individual element in the data set and by at least one variable parameter, and secondly at least one geometric measure associated with said variable geometric structure, which geometry may be arranged to measure a geometric property of a specific geometric structure relative to other specific such geometric structures, and in that the method further comprises the steps of a computer or several interconnected computers a) being caused to store the amount of data on a digital storage medium; b) for at least each element corresponding to a pixel in the image, and for at least one of the M element properties, is first caused to determine a specific geometric structure which can be obtained in relation to the element in question and by choosing for which parameter value or parameter values, specific Application textdoc 2010-0940 100172EN 10 15 20 25 30 geometric structure at least one of said geometric dimensions is maximum while it is geometrically limited by individual elements or groups of elements whose element property meets a certain predetermined criterion, and secondly is made to calculate at least one of said geometric dimensions for said specific geometric structure; c) in a database is stored an association between each element and. the for the element at least one calculated geometric dimension for each property for which a specific basic structure has been calculated in step b); d) is made to perform a main classification of the data set, which main classification is made to be based on a comparison measure between two elements' respective sets of associated geometric measures, each element being made to be associated with a certain class in a main set of classes; and e) on a digital storage medium is caused to store the image classified in step d) in the form of a set of pixels with the same dimensions as. the image, where each pixel is associated with the class in the main set of classes with which the element corresponding to the pixel in the image that said variable is associated in step d); and of geometric structure, dimension and criterion are not chosen so that the dimension calculated in step b) constitutes only a measure of the minimum distance between the element in question and the nearest element associated with the element property in question.
    The invention will now be described in detail, with reference to exemplary embodiments of the invention and the accompanying drawings, in which: Figure 1 is a first image in the form of a photographic, digital satellite image in gray scale over a partially built-up area; Application text doc 2010-0940 100172EN 10 15 20 25 30 Figure 2 shows a conventional, kernel-based segmentation in five classes for the satellite image illustrated in figure 1; Figure 3 shows a conventional, window-independent classification in twenty-nine classes for the satellite image illustrated in Figure 1; Figures 4a, 4b, 4c, 4d, 4e and 4f are schematic diagrams illustrating the determination of a specific geometric structure according to various preferred embodiments; Figure 5 is a graphical illustration of the distribution of the respective surface of a geometric structure determined according to Figure 4b for each pixel in the image shown in Figure 1 and for a particular specific of the classes illustrated in Figure 3; Figures 6a and 6b show a main classification according to the invention in two classes for the satellite image illustrated in Figure 1; monochrome image Fig. 7 ar * en. second image, in the form of one comprising two areas with different textures; Figure 8 shows a conventional K-means segmentation in. Five classes for the image illustrated in Figure 7; Figure 9 shows a conventional K-means segmentation in ten classes for the image illustrated in Figure 7; Figure 10 is a graphical illustration of the distribution of logarithmically scaled respective area of a geometric structure determined according to Figure 4b for each pixel in the image shown in Figure 7 and for a particular specific of the classes illustrated in Figure 9; Figure 11 shows a first main classification according to the invention in five classes for the satellite image illustrated in Figure 7; Figure 12 is a graphical illustration of the distribution of the respective surface of a geometric structure determined on a set similar to that of Figure 4b for each pixel in the image shown in Figure 7 and for a pixel. certain specific of the classes illustrated in Figure 9; Figure 13 shows a second main classification according to the invention in five classes for the satellite image illustrated in Figure 7; Figure 14 is a graphical illustration of the distribution of the respective surface of a geometric structure determined according to Figure 4a for each pixel in the image shown in Figure 7 and for a particular specific of the classes illustrated in Figure 9; Figure 15 shows a third main classification according to the invention in five classes for the satellite image illustrated in Figure 7; and Figure 16 is a flow chart illustrating a method of the present invention.
    Figure 1 is thus a photographic satellite image with a pixel resolution of 10x10 m over a certain partially built-up geographical area. The fact that the image is "photographic" means here that it is produced by light being captured and interpreted by an image sensor. In other words, the image can be several called produced by reading light over one or wavelength ranges. The image in Figure 1 is such a composite image, based on 5 different spectral bands which have then been converted into a gray scale in a conventional manner. The image is stored in digital form on a suitable storage medium in the form of a set of pixels, each pixel being associated with image information in at least one channel, in this case five channels, for light intensity. Figure 1 shows a Uàd fl ænæbnell image, where the pixels are arranged in a two-dimensional coordinate system, and where each pixel is associated with a single channel of light intensity (shade of gray).
    The conversion to grayscale is performed to increase clarity In the embodiments described below in the figure. as Application text doc 2010-0940 100172EN 10 15 20 25 30 uses the image illustrated in figure 1, the information was used in all five channels.
    It will be appreciated that even images with more than two dimensions, and with light intensity over more than two channels (e.g. light intensity information stored according to RGB standard), can be used according to the present invention.
    It will also be appreciated that the pixel values have been obtained by sampling the light radiation towards the image sensor. Similarly, other physical phenomena, such as sound, other types of radiation than electromagnetic, and. so on, sampled and presented spatially in the form of an image. This is the case, for example, with an ultrasound image.
    Furthermore, the image according to the invention is of such a format that it can be represented in the form of or as part of an N-dimensional data set comprising elements, in that at least two of the N dimensions constitute respective axes in the image, so that a certain sampled pixel in the image corresponds to a certain element in the data set.
    Thus, the data set can have the same dimensionality as the image, each dimension in the data set corresponding to a respective image axis. other It is also possible that the data set hnæfaü fi r dimensions than those that describe image axes. For example, it may comprise two or three geometric dimensions as well as at least one temporal dimension, the image consisting of a two- or three-dimensional, over at least one time dimension variable, image of an event. This is the case, for example, with the above-described image of a heart in three spatial dimensions as well as two temporal ones. Other types of additional dimensions are also conceivable, such as a temperature dimension. or * a dimension sonl describes some other continuous parameter value that gives rise to a spatial context that relates an element to its neighbors.
    Thus, an image can be a section of the data set, where the layer for the image in the data set along one or more dimensions is constant. It will be appreciated that one and the same three-dimensional set of data may include more than one image within the meaning of the invention, and in addition images of different dimensions. In addition, the amount of data can of course also include additional data, irrelevant to the method of the invention.
    According to the invention, furthermore, each element in the data set is associated with M numerical care, each of which represents a numerical measure of a property of the element in question, of which at least one property represents image information in a respective sampled channel as above.
    The data set thus has an outer dimensionality of N and an inner dimensionality of M.
    Thus, for example, the image may comprise a two-dimensional magnetic field in which each element comprises image information in at least one channel, each of which indicates light intensity for a certain wavelength spectrum. This year, for example, is the case with a two-dimensional satellite image. Another example is that the image is a three-dimensional image of a three-dimensional object, each element comprising image information in at least one channel indicating the care of a material property in the low in the object corresponding to elements in question. This is the case, for example, with a three-dimensional magnetic camera image.
    Application text doc 2010-0940 100172EN 10 15 20 25 30 lO The M numerical values can represent different types of data.
    Sampled channels may include, for example, one or more channels, such as. grayscale data. in a channel, which. then constitutes a numerical value per element, or RGB-coded color values in three channels, which then constitute three numerical values per element.
    In addition to sampled image information, for each element, one or more of the M properties may also represent spatially located, predetermined, or calculated additional information. Preferred examples of such spatially located information include information about in areas; calculated or manually and / or image occurring edges; identified points of interest; data retrieved from an external database that has been mapped so that it has the same dimensions as the image, whereby spatially located information can be attributed to individual elements.
    Edges, areas, etc. can be calculated in advance using conventional image processing algorithms or filters such as the Harris corner detector and the Sobel operator.
    Points of interest can, for example, be marked by a user, such as your doctor smn marking a specific point that represents an area of interest in an X-ray.
    Alternatively, different types of additional spatially oriented data can be retrieved from external databases. Examples include the presence of grocery stores or other known points of interest in a satellite image; meteorological data for the land areas represented in a satellite image; local measurement values for geographical data such as slope, altitude and land distribution; demographic data such as population density and land ownership; and the existence of an Application Text Doc 2010-0940 100172EN 10 15 20 25 30 ll certain; spelztral fl nix i. a nuiltispectral satellite image and the presence of a certain color in an image.
    In order to be represented in the form of one or more of the M element properties in the data set, such additional spatially oriented data are mapped by being transmitted to a raster * with the same dimensions as. the set of data, after which the resulting numerical values are entered as one or more of the element properties of each element.
    It is also preferred that one or more of the M element properties carry numerical information regarding class affiliation in a set of classes used in an initial classification of the elements included in the data set. In this case, for example, one element property per eleamerite can be Ined. with the help of 'ett; nlimeric value indicate the class affiliation of the element in question. Alternatively, one element property per element and class in said set of khmser in the embedded classification may indicate class affiliation to the class in question for the element in question. In this. case. For example, the class affiliation may be binary, the element in question either belonging to the class or not, or represented on a sliding scale by a number between 0 and 1, where 0 means no association between the element in question and the class in question and where l means full association. A preferred such initial classification is described below in more detail.
    It is preferred that the M element properties be coded as friendly. The set of a certain form scalar, numerical element property for all elements thus a scalar coded information layer that extends over the entire natural set of data. In general, the values can be binary or real, positive and / or negative, depending on what they are arranged to represent.
    Element properties can further represent probabilities.
    It is also obvious that the M information layers can be mutually overlapping in the sense that several properties for one and the same element can, for example, be zero different.
    It is desirable to provide an automatic classification of such images, in order to distinguish in a simple manner and with high accuracy. different types of phenomena in the picture, such as roads, buildings, forests, arable land, parks, etc.
    The image shown in Figure 2 is the same image as that in Figure 1, but where the image has been processed through. en] mmvenUbnelL kernel-based segmentation. A window consisting of 5x5 pixels has been swept over the image and for each pixel the standard deviation has been calculated with respect to the light intensity of the pixel in question. These metrics have since formed the basis of a conventional class division based on. on threshold values, which has resulted in the illustrated classification. The result is illustrated in Figure 2, where different classes are illustrated with the help of different shades of gray. It is clear from the figure that such a method is successful for finding different classes and for assigning for each pixel an association with a certain specific class. In addition, many of the pixels that constitute different types of terrain, such as "road", in the image shown in Figure 1 are grouped in that they are associated with one and the same class. However, there is a relatively strong noise in the classified image.
    Too many pixels have been associated with a different class than most pixels that make up the same image type in the original application text 2010-0940 100172EN 10 15 20 25 30 l3 image. Therefore, it is difficult to perform, for example, an automatic mapping based on such a classification. Figure 2, a Figure 3 shows, in a manner similar to that of the classification of the image shown in Figure 1, with the difference that the classification in Figure 3 is performed in accordance with a conventional, window-independent cluster analysis of K-mean type over 29 classes. As is clear from Figure 3, this classification has the same types of problems as that of Figure 2. A flow chart illustrates a Figure 16 as a method in accordance with the present invention.
    Thus, in accordance with the invention, in a first step 101, a computer, or several interconnected computers, are caused to store the amount of data digitally on a digital storage medium in the form of a set of elements having a structure as described above.
    According to the invention, for the image to be classified or for a certain category of such images, a certain predetermined, variable geometric structure is identified. The variable structure is arranged to be able to extend in at least two of the N dimensions in the data set, and in other words can have an extension in at least these two dimensions. It is noted that the variable structure itself may be of a dimensionality lower than the number of dimensions in the data set that it may extend over, such as, for example, a one-dimensional line running from one image pixel to another in a two-dimensional image represented by the data set. .
    That the variable geometric figure is to be "variable" is interpreted as meaning that its shape can be determined unambiguously by the value of at least one variable parameter. Examples of such parameters are the radius of a circle or sphere, or the radii of an ellipse; the distance of a segment from a midpoint; an extension angle of a segment; a center or center of gravity of the geometric figure; a function that describes the periphery of the figure relative to a fixed point; and so on.
    In addition, the extent of the geometric figure in relation to an individual element in the data set is determined, in that 1neni11ger1 that; the different element constitutes a fixed reference point in relation to the variable geometric figure. According to a preferred embodiment, the individual element in question is always included or at least completely surrounded by the variable structure in the section of the amount of data in which the shape of the structure can vary.
    It is preferred that the identified variable geometric structure be used for all elements treated in a method according to the present invention. This minimizes the amount of manual work that must be performed.
    When a variable geometric structure is fixed, by selecting a certain element and by selecting the variable parameter values to specific values, a specific geometric structure is produced, which in this case thus describes an unambiguous geometric shape in the plane, space or the like.
    Furthermore, for the image to be classified or for a certain category of such images, at least one geometric dimension associated with the variable geometric structure is determined. The measure is arranged to network the geonuatric property of a certain specific geometric structure in relation to other specific such geometric structures. In other words, the measure is arranged to measure the ratio of the fixed instance to the geometric property of a certain specific, of the variable geometric structure in other specific such instances of the same variable geometric structure, for which other elements and / or other parameter values have been selected. With the help of such a measure, different specific structures can thus be compared geometrically.
    According to a preferred embodiment, the measure is invariant in terms of pixel resolution of the image, i.e. invariant, up to numerical precision, with respect to the sampling of the image. In addition, it is preferred that the measure is invariant during rotation, and furthermore that the measure, or at least the ratio between two measured values measured with the same, is invariant during scaling of the image. In the case of geometric measurements, in other words, the measured value for a specific structure * does not change in relation to the corresponding measured values for other specific structures as a function of scaling and / or rotations of the image. Preferred examples of invariant dimensions are maximum radius, length, circumference, area, surface area, degree of convexity, central moment, center of gravity and / or degree of circularity or sphericity of the structure.
    It is further preferred that the dimensions used are arranged to measure some aspect which is attributable to the geometric size of the geometric structure.
    In a second step 102, the computer or computers are then caused, for at least each element corresponding to one pixel in the image but preferably for all elements in the data set, and for at least one of the M element properties, preferably for all M element properties, to first determine a specific geometric structure that can be obtained in relation to the element in question and by selecting parameter value or Application text doc 2010-0940 100172EN 10 15 20 25 30 16 parameter values. The specific geometric structure determined for a particular element is arranged to maximize at least one predetermined geometric dimension of the type described above, while the structure is geometrically limited by elements present in the data set whose element property meets a certain predetermined criterion.
    Figures 4a-4f illustrate in principle and simplified different ways of determining such specific geometric structures on the basis of a certain individual element and a certain element property among. the M properties. For the sake of clarity, all the examples illustrated in Figures 4a-4f relate to two-dimensional images. It will be appreciated, however, that for images of higher dimensionality, and possibly also including continuous dimensions which do not in themselves constitute image axes, the same principles are used to provide analogous methods for determining specific geometric structures. For example, instead of the circular structure illustrated in Figure 4a, a three-dimensional sphere can be used, which is allowed to expand to elements with certain properties. Correspondingly, a set of three-dimensional, cone-like parts can be used instead of the cake piece-like parts in Figure 4b.
    According to a first preferred embodiment, which is illustrated in Figures 4a-4c, it is variable. geometric structure such that it can. an expansion rule is defined by means of which the geometric structure can expand stepwise on the basis of an individual element in relation to which the extent of the geometric structure is determined.
    In Figure 4a, the geometric structure thus consists of a circle 13, which is defined by the fact that it can gradually expand from an individual element 12. In the example in question, the expansion takes place so that the circle has such a large radius 14 as possible, provided that said individual elements 12 are always included in the circle 13 without necessarily being in the center 15 of the circle 13. The expansion of the circle 13 is limited by allowing it to expand only until it reaches somewhere along its periphery one or more elements ll whose element property in question meets a certain criterion, for example that the element property is zero-separated or exceeds a certain predetermined value. For simplicity, only elements ll that meet this criterion are illustrated. Figure 4a illustrates the circle 13 for which the radius of the circle is maximum. In other words, the maximized geometric dimension in this exemplary case is the radius of the structure.
    More generally, the geometric structure comprises a closed geometric figure, which may be a circle but which may also be any other well-defined, closed structure. The geometric figure has a certain general shape which, according to a predetermined expansion rule, can gradually expand by growing while maintaining the certain general shape, the geometric extent of the structure being limited in the dimensions of the data set by allowing the shape to expand only to somewhere along its periphery. reaches individual elements or groups of elements whose element properties meet a certain predetermined criterion.
    It is preferred that the mold expands so that said individual elements are always a subset of the structure, which is exemplified above in that the element 12 is always present within the circle 13.
    According to a preferred embodiment, which is not exemplified in figure 4a but instead in figure 4b, næm exp a Applicable textdoc 2010-0940 100172EN 10 15 20 25 30 l8 exposes geometric shape in that all points along its periphery move straight out from said individual elements. It is in 1nånga.1: applications that the center of the structure is not limited to the element in relation to which the structure is determined, but that the individual element must instead only be a subset of the structure, since this gives a more even answer.
    Figure 4b illustrates, in a manner similar to that of Figure 4a, a second example of the determination of a specific geometric structure. The type of expanding geometric structure illustrated in Figure 4b is conventionally referred to as the "viewshed".
    The structure in this case comprises a collection of radial parts 23, which relation to each covers a certain angular range in the individual element 22 in question.
    The angular ranges are preferably the same size, but can also be irregularly distributed over an entire revolution. It is preferred to use at least 4 angular ranges, to ensure satisfactory results. According to a particularly preferred embodiment, as shown in Figure 4b, 8 equally large angular sections are used. This means that the calculations are especially simplified, which means low processor requirements for the computer or computers. According to another preferred embodiment, a larger number of equal angle sections are used, such as at least 50, more preferably at least 100 angle sections. This gives high precision and thus even end results.
    According to the expansion rule, each such radial part 23 can expand in radial direction independently of the other radial parts, but the expansion for each part 23 is limited by allowing its periphery to expand only up to element 21 whose element properties meet said criterion.
    In the examples illustrated in Figures 4a and 4b, the geometric structure is limited in that it is not allowed to include elements that meet a binary criterion, such as being birir: associated with the elemaüßgenäæp in question. In other words, for each element independent of other elements, it can be determined whether the element meets the said criterion or not.
    Fig. 4c, which is similar to Figure 4b, illustrates a preferred embodiment where the criterion is not of said binary type. In the same way as in Figure 4b, a number of angular sections 33 are allowed to expand from the element 32 in relation to which the specific structure is determined, and the radial expansion of the angular sections 33 is limited by elements 31. Only elements whose element property is zero separated are illustrated, for clarity. .
    In contrast to Figure 4b, on the other hand, at least a part of the variable geometric structure, in the illustrated example, at least one angular section 33 is limited in that the part is not allowed to include a number of elements 35 whose total association with the property in question exceeds a certain predetermined value. . Determination of the specific structure can thus be performed by allowing each angular section 33 to expand radially from the element 32, and each time it reaches another element 31, the total association between the elements 35 reached so far and the element property in question is calculated. In the event that the total association exceeds a certain value, the expansion of the angular section stops. at least a portion 33 of that portion 33 More generally, the geometric structure is limited by not being allowed to include a plurality of elements 35 which together satisfy a predetermined criterion of the association of the elements with the property in question 2010-0940 100172EN 10 15 20 25 30 20.
    In both Figure 4b and Figure 4c, the expansion of each angular section 33 is further limited by a maximum radial expansion 34. This is distance 24, in order to avoid extreme values. preferred but not necessary.
    Figure 4d illustrates, in a manner similar to that of Figures 4a-4c, an alternative preferred approach for determining a specific geometric structure, in relation to a particular element 42 and taking into account other elements 41. In this case, the geometric structure comprises a closed geometric figure * 43 with variable shape This means that the specific shape of the structure itself is not invariant, but can change depending on the location of surrounding elements 41 in the local vicinity of the element 42, based on the values of the parameters that determine the specific shape. Such a structure is preferred because it provides particularly even responses.
    In addition to the fact that, as described above, the shape 43 is limited by the elements 41 which it reaches and which meet a certain criterion regarding the element property in question or alternatively regarding the total element properties of several elements included in the mold 43, the spread 43 of the mold 43 is limited by a predetermined energy functional 43. Such an energy functional can be any suitable functional, such as that the surface or circumference of the mold 43 is kept constant or must not exceed a certain predetermined value.
    An example where it is appropriate to limit the surface of the mold so that it is not allowed to expand above a predetermined value is where the geometric structure is expanded from and in relation to a certain element like a balloon being inflated. up. In this and other cases, the structure according to the expansion rule expands by, at any given stage of expansion, only sections of the periphery of the structure whose expansion is not currently limited by individuals or collections of several elements meeting the said criterion are allowed to expand. In other words, the other elements operate in the local environment. to the element that. the expansion takes place in. relation to, as fixed stops for the expansion. Such a structure enables an even response even in the case of non-convex shapes of limiting elements in the local environment to the element in question.
    Figures 4e and 4f show, in contrast to Figures 4a-4d, an exemplary image in two dimensions, where the color of dark elements 51 and light elements 52, respectively, represents a binary value which, for example, indicates the element in question is associated with a certain class in an initial set of classes, alternatively with a certain other element property.
    Figure 4d illustrates a geometric structure in the form of a straight line 54, which starts from the element 53 in relation to which the structure is determined and runs straight out from this element. The specific geometric structure is determined as the structure that maximizes the dimension that constitutes the length of the line 54. In this case, the specific structure is thus not determined by expanding the structure stepwise in relation to the element 53. Instead, a conventional and appropriate algorithm is used to effectively determining the longest line that can be written into the data set, which starts in the element 53 and whose path is limited to element 52 with a certain binary value with respect to the element property in question.
    Figure 4d illustrates the longest possible such line 54 which is limited to a distance over light elements 52. Such a structure allows calculations with small claims to processing power.
    More generally, it can be said that the variable geometric structure comprises an elongate part, the extent of which in a longitudinal direction is limited by not allowing it to cross elements whose element properties meet a criterion.
    The structure illustrated in Figure 4e thus consists of a per se one-dimensional line which extends over the two-dimensional image surface. Figure 4f illustrates the same amount of data including dark 61 and light 62 elements, but an alternative elongate structure 64 extending from a particular element 63 in relation to which the structure 64 is determined.
    As is clear in Figure 4f, in this case the elongate part 64 is not straight, but its extent is allowed to describe a curved curve through the image. In addition, the curve 64 is not in itself one-dimensional, but constitutes a two-dimensional body with a certain width 65. It will be appreciated that in, for example, a three-dimensional image. can. a three-dimensional, elongated body is used in an analogous manner. Such a structure may, for example, be useful for classifying veins in medical images of the human body or in other types of images where elongate, hollow structures are relevant.
    The specific geometric structure is determined in this case as the elongate body 64, with a constant width 65, which is the longest possible, provided that the elongation of the body 64 from the element 63 is described by a predetermined Application Text Document 2010-0940 100172EN 10 15 20 25 30 23 algorithm and without touching any dark elements 61 in the image.
    It will be appreciated that a straight structure of the type illustrated in Figure 4e may also be provided with a certain width as is the case in the structure of Figure 4f. In this case, the structure according to Figure 4e thus becomes in itself two-dimensional rather than one-dimensional.
    For: the amnira, the computer or the interconnected computers are caused to calculate at least one geometric dimension of the specific geometric structure as determined above. the calculated dimension may be the same as or different from the dimension maximized by the particular structure itself, but the dimension calculated for the specific structure in step 102, which is then stored in step IO3, is of the same general type as the dimension which is maximized in step 102 of the specific structure. Where applicable, the above applies to the maximized Inatt: äver1 for the calculated and stored measure.
    In addition. it is preferred that more than one, preferably at least three different, and preferably also invariant, such geometric dimensions be calculated for each particular specific structure. save risk This allows more information to be extracted from the calculated specific structures for all pixels.
    Non-limiting examples of suitable dimensions to be calculated for Figures 4af4d include the relative position of the W W rea and the center of gravity of the structure in relation to the element in relation to which the specific structure was determined; for Figure 4d the degree of circularity and the degree Application Convex 2010-0940 100172EN 10 15 20 25 30 24 of convexity; for Figures 4e-4f the length; and for Figure 4f the degree of crookedness.
    The same variable geometric structure and the same at least one geometric dimension are used for all elements and all treated element properties in step 102.
    In step 103, for all elements processed in step 102, the computer or the interconnected computers are then stored in a database to store an association between each element and that of the element at least one calculated geometric dimension for each property for which a specific geometric structure has been calculated. in step 102. Database may be the operating memory of the operating computer, a file on a hard disk, or another type of internal or external database.
    Thus, for each element in the image, specific geometric structures for several different properties are preferably determined, and one of these different structures is calculated and for each several different geometric dimensions. Associations between each element and all these dimensions are thus stored in the database, so that each element in the database is typically associated with M * x dimensions, where x is the number of calculated dimensions per specific structure.
    It is preferred that an association between each element and at least two calculated geometric dimensions be stored in the database. These dimensions can thus relate to specific geometric structures with a / seeruïe of. <> Equal element properties and / or to the same specific geometric structure.
    It is especially preferred that a specific geometric structure be determined for each element and for at least two of the M element properties.
    It will be appreciated that step 103 may be performed immediately following step 102 for each individual element, or that step 103 may be performed for all elements in a sequence after step 102 has been performed for all elements in a sequence, 102 or that steps and 103 are performed otherwise, as long as both step 102 and step 103 are performed for all elements in the image.
    As described above, it is preferred that at least one of the M element properties is initially populated with data representing association between the element in question and a particular class or a number of certain classes of a first set of classes in a first, initial classification of the data set.
    It is preferred that the computer or the interconnected computers in this case be caused to initially perform such a first classification of the digitally stored data set, and store the result of this classification in a first database. In the present exemplary embodiment, the classification shown in Figure 3 is stored in the first database. Herein, the result of this classification is referred to as "the first classification". The first database may be the same as the database described above or another. In the latter case, the first database may be of the type described above.
    The first classification can be a cluster analysis of K-mean type, or any suitable kernel-based or window-independent classification, as long as the classification is based on each respective pixel's stored image information and possibly also on other already stored Application textdoc 2010-0940 100172EN 10 15 20 25 30 26 properties for each respective element. The first classification can be based on predefined, static classes, a predefined number of but variable classes, or the classes and their definition can take shape during the course of the classification.
    The first classification can also be a supervised or an unsupervised classification. A supervised classification consists of two steps. In the first step, representative areas are identified in a representative image, and a numerical description of the properties of the representative areas is provided. In the second step, the numerical descriptions of representative areas are used as classes in a final classification of an image. In an unattended classification, each pixel in a certain image is associated with a certain class, where the properties of all classes are determined during the process, ie. without prior knowledge of the types of image areas that may occur. According to the invention, it is preferred to use an unsupervised classification, since such a classification can operate over a wider range of different images without special adaptations.
    Examples of useful monitored classifications are "Minimum-Distance-to-Means classifier", "Parallelepiped Classifier", "Gaussian Maximum Likelihood Classifier", "K-nearest neighbor" and "Maximum Likelihood classification".
    Examples of useful unsupervised classifications are various hierarchical clustering methods, partition clustering methods, “K-mean” clustering and “Self organizing maps”.
    Figure 5 graphically illustrates the calculated geometric dimension "area" for a respective specific geometric structure determined in the manner illustrated in Figure 4b, with eight equally distributed angle segments which are expanded out from the element in question and limited of other elements associated with a particular individual class in the classification illustrated in Figure 3. The calculated area of each specific structure is illustrated using grayscales - the more accurate: the color of the element in Figure 5, the larger the area of the specific structure determined in relation to the element in question.
    Black elements represent elements with a maximum allowable surface area, larger than which structures have not been allowed to expand.
    According to the invention, the computer or computers then perform, in a fourth step 104, a main classification of the amount of data, this time based on the data stored in the second database and preferably for and based on all elements corresponding to pixels in the original image.
    According to the invention, the metric relevant for this second classification for each element is a comparative measure between two elements' respective sets of associated saving risk som calculated in step 102, and the main classification is thus based on this comparative measure. These calculated dimensions can thus consist of one or more measured values per element, which is why it may be necessary to perform a multidimensional cluster analysis as part of the second classification.
    Examples of useful comparative measures include the usual Euclidean distance, possibly with a scale factor. Non-scalar, inherently conventional, comparative measures can be used to compare two vectors comprising several different geometric measures, and so on. Furthermore, in the case of vectors, an isotropic or anisotropic scaling can be performed on the measured value vectors before they are compared, such as by means of a matrix operation.
    Application textdoc 2010-0940 100172EN 10 15 20 25 30 28 Suitable types of classifications for that main classification. are the sonl above stated as. suitable for the first classification, and the main classification may be of the same type as, or different from, the first classification. However, it is preferred that the main classification be of an unsupervised type. In addition, it is preferred that the main classification be of an unsupervised one regardless of whether any first type classification is of a supervised or unsupervised type. Such a procedure allows automatic and precise classification over a wide range of different original images.
    The main classification can also include a manual segmentation of the data stored in the database, so-called using a "region-growing" technology, where a user manually selects a starting point ("seed").
    The outcome of the main classification is thus an association between each respective pixel in the original image and a certain specific class of a main set of classes which are either defined in advance or which are defined while the main classification is being performed.
    Based on the comparison measure described above, it is also preferable to, before the main classification, reduce the amount of noise by averaging elements for which the calculated geometric measures correspond. This results in a filtering, or more generally regression, of the amount of data, and can serve to improve the result of the main classification.
    Thereafter, in a fifth step 105, the computer or the interconnected computers according to the invention store on a digital storage medium the image classified in step 104 in the form of a set of pixels, each respective pixel being associated. with the class in the main set of classes with which the element corresponding to the pixel in the image is associated in step 105 as described above. The digital storage medium is the same as the one on which the original image was stored or another, and the storage technique may be as described above with respect to the original image. What is important is that the set of pixels, where * and. one with. an association to a specific class, is stored in a way that allows this information to be represented as a classified image, where the position of each pixel in the classified image. corresponds to the position of the element in the original image, and where the value of the pixel, such as color intensity in one or more channels, corresponds to the class with which the pixel in question is associated. It follows that the classified image has the same dimensions in terms of the number of pixels as the original image. In subsequent steps, of course, the resolution of the classified image can then be adjusted.
    Figure 6a, and Figure 6b in detail, illustrate one as definitively classified. image, based. on the original image shown in Figure 1 and performed over ten different main classes. To increase clarity. in the figure. these ten classes have been merged into two different collection classes, which two different classes are represented by different grayscales. Furthermore, the classified image is illustrated against the background of a high-resolution aerial photograph of the geographical area illustrated in Figure 1, so that the distribution of the classes is clearly visible. The classification illustrated in Figures 6a and 6b is thus based on the largest area of the respective geometric structures according to Figure 4b for each pixel in the original image and with respect to each of the classes that resulted. from the first classification illustrated in Figure 3. illustrates an image other than the one illustrated in Figure 7 in Figure 7. As can be seen from Figure 7, the image comprises a central area with a different type of texture than the rest of the image.
    Figure 8 illustrates a naive segmentation of the image in Figure 7 using K-means in 5 classes based on the gray value of the respective pixels in the image. Furthermore, Figure 9 is a segmentation of the image in Figure 7 using K-means in 10 different classes. It is noted that the separation of the two areas with different textures is deficient in both Figure 8 and Figure 9. Not even a manual challenge merging classes can give rise to a low noise division into areas with different textures.
    A two-dimensional data set was put together: in which the pixels of the two-dimensional image each represented an element of the data set, and where a ten-dimensional property vector comprising the affiliation of an element to each of the ten classes from the classification illustrated in Figure 9 was associated. down. where the elements of the data set.
    A variable geometric structure was then defined as a viewshed. with four equally distributed angular segments that were allowed to expand independently up to the first element that was associated with a certain property, in other words a certain class. Specific geometric structures were determined for each element, which specific geometric structures maximized the geometric measure "total area for all four expanded parts". A geometric measure in the form of the logarithm of the total area of the structure was calculated for each specific geometric structure. Figure 10 illustrates the measure thus calculated with respect to a certain of the classes illustrated in Figure 9, where black color means large or a predetermined maximum area and successively lighter shades of gray mean ever smaller areas.
    A respective association between each element in the image and each of the ten calculated dimensions per element was stored in a database, after which a main classification was performed based on this data. The main classification consisted of a K-mean classification of over five. classes. Figure 11 illustrates the result. As is clear from Figure 11, this algorithm distinguishes the different areas with different types of texture, with relatively low noise. It is noted that some disturbing artifacts are present in the classified image. as a consequence of the fact that only four expansion directions were used in the determination of specific geometric structures.
    Figure 12 illustrates, in a manner similar to that of Figure 10, a calculated geometric dimension with respect to some of the classes illustrated in Figure 9, but where a different variable geometric structure has been used. This structure included thirty-six one-dimensional straight lines that were allowed to expand independently in evenly spaced directions until they struck an element associated with a particular and of the classes illustrated in Figure 9. The maximized geometric dimension was the length of each of the lines. The calculated geometric dimension of the specific geometric structure was the total area of the two-dimensional polygon that was stretched using the outer endpoints of the expanded lines.
    Application text doc 2010-0940 100172EN 10 15 20 25 30 32 Figure 13 shows the result after a main classification, over 5 classes using K-means, of the result after calculation of such geometric dimensions for all elements and all ten classes shown in figure 9. As is clear, the result was satisfactory, especially if the two classes describing the inner texture were merged.
    It is noted that a similar end result was obtained if a viewshed with 360 equally distributed angular parts was chosen as a variable geometric structure, indicating that thirty-six different directions of expansion are sufficient to achieve good results with the type of images exemplified by Figure 7. It has been found that for many applications, about ten different, evenly distributed directions or more are sufficient to achieve good results. If about ten directions are used, the requirements for computing power are limited at the same time.
    Figures 14 and 15 are similar to Figures 10 and 11 and 12 and 13, respectively, but here, as a variable geometric structure, expanding, non-centered circles of the type described above in connection with Figure 4a were chosen instead. From the final result in Figure 15, it is clear that this method according to the present invention provides a very good separation of the two different types of texture in the original image in Figure 7.
    A method according to the present invention, wherein an organization of image and other data in a data set of the type described above comprising elements with respective element properties, is followed by a calculation of at least one respective geometric dimension for a respective specific geometric structure for each respective element in the amount of data and with respect to one or more element properties, and then followed by a main classification Application text doc 2010-0940 100172EN 10 15 20 25 30 33 based on these calculated dimensions, solves the problems described above.
    Depending on the specific application, the geometric dimensions can be chosen so that they serve to estimate the density of a certain element property in a local environment for a certain element. The main classification can therefore achieve very sharp class boundaries. This is an effect of the fact that the variable geometric structures can be selected so that the specific structures of two adjacent elements are often very similar or exactly coincide, giving almost consistent eclipse values in piecewise areas in the finally classified image.
    As an illustrative example, it is conceivable that two exemplary geometric structures of the type illustrated in Figure * 4b are 'surrounded' by a certain element property, which is represented by an element in a ring around the two elements in relation to which the structures expand. . If this area is approximately convex, the expanding sectors will cover the same convex area. This area for all elements inside leads to the fact that, apart from numerical effects such as only a finite number of sectors being expanded from each pixel, a measured value such as the total sector area will be the same for all elements in the area. Thus, all the pixels in this area will also be classified in a similar way in the main classification step 104.
    The exact choices of variable geometric structure, type of dimension to be maximized in a specific structure, type of limitation with respect to the properties of surrounding elements and type of dimension to be calculated depend on the current from case to purpose and conditions, and determined Application textdoc 2010 -0940 100172SE 10 15 20 25 30 34 fall. However, the principles described herein remain the same regardless of these choices. Thus, the probability that nearby pixels are associated with the same final class will be relatively large. In a similar way, the probability is small that a pixel, which is similar to a group of pixels which in the final classification falls under a certain class but which is isolated from these other pixels in the image, falls under the same class. In other words, a method according to the present invention has a strong tendency to group adjacent pixels in the same class and thereby create contiguous areas of pixels associated with one and the same class. These effects cause low fragmentation in the pre-classified image. On the other hand, the calculated geometric dimensions will change rapidly between adjacent pixels in case the element properties change rapidly locally in an area of the image. This means that a method in accordance with the present invention will in a relatively precise way produce a final classification where boundaries between the final classes really represent relevant boundaries and transitions in the original image.
    A method according to the present invention can be designed so that it gives better results than a method according to the above-mentioned Swedish patent application 0950140-4, since a geometric measure according to the above generally constitutes a more stable and noise-free estimate of the local property density than only a minimum distance to a certain class.
    Taken together, the method described above can be used to automatically and without extensive prior knowledge of a particular application text 2010-0940 100172EN 10 15 20 25 30 35 image content obtain a relevant classification with low splitting. This is particularly the case. According to a preferred embodiment, in step 102, for each element, at least one geometric dimension is calculated for a specific geometric structure with respect to each of all element properties in the data set, and all calculated dimensions are also used in this case in the main classification. This maximizes the utilization of the information provided. can be found in the original image and any additional spatially related information.
    According to a preferred embodiment, the first classification comprises a classification using a K-mean cluster analysis followed by a classification using a maximum likelihood cluster analysis, in which the result from the K-mean classification is used as a starting point. According to a further preferred embodiment, the main classification comprises er1 K-mean cluster analysis. The present inventors have discovered that such a process gives a good end result.
    The present inventors have further achieved good results in the case where the number of classes in the first classification was either set in advance between 20 and 100, or when the first classification worked with a variable number of classes and where the classification in this case was adjusted so that the resulting number of classes became between 20 and 100.
    According to a particularly preferred embodiment, the so-called Akaike information criterion is used to determine an optimal number of classes in the first classification by balancing the complexity of the model against the intraclass variance. an Akaike- In other words, the number of classes is selected so that the information criterion is maximized for the classification at Application textdoc 2010-0940 100172EN 10 15 20 25 30 36 this choice of number of classes. See Akaike H., "A new look at the statistical model identification", IEEE Trans. Automat.
    Contr. AC-19: 7l6-23, 1974 [Institute of Statistical Mathematics, Minato-ku, Tokyo, Japan].
    Correspondingly, it has been found that a suitable number of classes in the main classification is between 5 and 20, especially when a first classification is performed and the number of classes in this is almost 20 and 100. Depending on the detailed application, such a number of classes generally gives a definitively classified image with relevant classification 'of individual pixels and. with. limited noise in most classes. The choice of the number of classes in the main classification. depends 'partly on' which. type of results desired from the process if according to the present invention, partly on further processing of the obtained classification is to be performed in later steps. It is especially preferred to use the Akaike criterion to determine the number of classes also in the main classification, in a corresponding manner as described above.
    All scalar quantities handled in the different steps of the process can be further transformed by a non-linear transformation, 2/2 so sonl x or X1, before being used in calculations or comparisons. This can give relatively reduced weight to, for example, very long structures, if the square root from the length measure is used.
    As discussed above, a method according to the present invention can be advantageously used to automatically obtain a relevant classification of two-dimensional images, especially for photographic images. Classification of such photographic images containing image information in a channel, such as grayscale light intensity, has been found to give good results.
    A method according to the present invention is also very useful in the case of photographic images with image information in three or more channels, such as for example color intensity information according to some conventional color scheme with three channels such as RGB, Lab, YUV, HSV or NCS, or a representation where several different channels represent color intensity over different wavelength ranges.
    The latter uses the type of color information models, for example in satellite imaging, and then often with four or more channels for different wavelengths.
    From such a two-dimensional photographic satellite image, a method according to the present invention can thus be used, for example, as an automatic tool in mapping, demographic or geographical surveys, environmental analyzes, etc. In the case of other types of two-dimensional photographic images, a method according to the present invention can be used as part in systems for, for example, computer vision, visual inspection of manufactured details or of input or raw materials, auxiliary systems for photography to automatically identify various objects and / or find suitable focus points, digital image processing, remote reading, microscopy, digital image analysis and processing and so on.
    It will also be appreciated that a method of the present invention is useful for automatically classifying three-dimensional images, in particular, and as discussed above, in the case of three-dimensional images representing the tissue structure of humans and animals in and veterinary fields. Such three-dimensional medical images are in fact three-dimensional images of three-dimensional objects, where each pixel includes image information indicating the value of one or more material properties in the position in the object to which the aff pixel fl i is applied. in question. The material properties can, for example, be material density and / or the presence. of a marker substance. It is common for such images to have image information only in one channel. In particular, similarly segmented three-dimensional images can be used to advantage to analyze the human body, to support surgeons undergoing surgery and to analyze three-dimensional images in the micro- and nanotechnology areas. Particularly in medical applications, it is useful to apply a method of the present invention to an image comprising one or more temporal dimensions as described above.
    According to a preferred embodiment, the computer or computers further initially perform a kernel-based processing of the original image, and the result from this is then appended to the data set as one of the element properties of all elements. An example of this is that a value for each pixel is calculated for the texture in a limited area around the pixel in question, for example a 9x9 matrix with the pixel in question in the center, and that this calculated value is then computed to constitute one of the M element properties of the element in question.
    In this way, in some applications the accuracy of the final classification can be increased.
    According to the invention, the variable geometric structure, the geometric dimension and the criterion of element properties are not used which are used to determine how a structure is limited so that the geometric dimension calculated in step l02 only constitutes a measure of the smallest distance between the element in question and the nearest element. which is associated with the element property in question.
    Application text doc 2010-0940 100172EN 10 15 20 39 Preferred embodiments have been described above. However, it will be apparent to those skilled in the art that many changes may be made described departing from the spirit of the embodiments of the invention.
    For example, the present process can be performed iteratively. In other words, after a first classification, calculation of geometric dimensions and a main classification, a further calculation can be performed by geometric measurements for specific geometric structures determined with respect to one or more of the classes resulting from said main classification, and thereafter further a main classification is performed based on these calculated measurements.
    The latter calculation of dimensions and the classification can be varied within the framework of what is stated above for such a calculation and for the main classification. For some applications, such a classification in several iterative steps can give even better results.
    Thus, the invention should not be limited by the described embodiments, but may be varied within the scope of the appended claims.
    Application text doc 2010-0940 100172EN
    权利要求:
    Claims (16)
    [1]
    A method for classifying a two-dimensional high-dimensional image, wherein an N-dimensional data set comprising elements is made to represent the image by at least two of the N dimensions constituting respective axes in the image, so that a certain pixel in the image corresponds to a certain element in the data set, and in that each element is associated with M numerical values each representing a measure of a property of the element in question, of which at least one property represents image information in a respective sampled channel, characterized in that the method comprises identifying, first, a certain predetermined, variable geometric structure, the extent of which in at least two of the N dimensions of the data set is determined in relation to a single element in the data set and of at least one variable parameter, and for that æm a at least one geometric dimension associated with said variable geometric structure, which geometric measures are caused to be arranged to measure a geometric property of a specific geometric structure in relation to other specific geometric structures, and in that the method further comprises the steps of bringing a computer or several interconnected computers a) to storing the amount of data (10l) on a digital storage medium; b) for at least each element corresponding to a pixel in the image, and for at least one of the M element properties, is first brought to determine a specific geometric structure (13; 23; 33; 43; 54; 64) which can be obtained in relation to the element (l2; 22; 32; 42; 53; 63) in question and by selecting parameter value or parameter values, for which specific geometric structure at least one of the geometric dimensions mentioned in Application text doc 2010-0940 100172EN 10 15 20 25 30 41 is maximum at the same time as it is geometrically limited by individual elements or groups of elements whose element property in question meets a certain predetermined criterion, and secondly is caused to calculate at least one of said geometric dimensions for said specific geometric structure (102); c) storing in a database an association between each element and the element at least one calculated geometric measure for each property for which a specific geometric structure was calculated in step b) (103); d) is made to perform a main classification of the data set, which main classification> is made to be based on a comparison measure between two elements' associated geometric and sets of dimensions, each element being made to be associated with a certain class in a main set of classes (104); and e) on a digital storage medium is caused to store the image classified in step d) in the form of a set of pixels with the same dimensions as the image, each pixel being associated with the class in the main set of classes with which the element corresponding to the pixel. in the image associated with step d) (105); dimensions and and. of that said * variable. geometric structure, criterion is not chosen so that the dimension calculated in step b) constitutes only a measure of the minimum distance between the element in question and the nearest element associated with the element property in question. Application text doc 2010-0940 100172EN 10 15 20 25 30 42
    [2]
    Method according to claim 1, characterized in that the variable geometric dimension is invariant during scaling of the image.
    [3]
    A method according to claim II or 2, characterized in that, in step c), an association between each element and at least two. calculated geometric dimensions are stored in the database.
    [4]
    4. A method according to claim 3, characterized in that, in step b), a specific geometric structure is determined for each element and for at least two of the M element properties.
    [5]
    Method according to one of the preceding claims, characterized in that the variable geometric (13:23; 33) of the structure is caused to be defined by an expansion rule by means of which the geometric structure can expand stepwise on the basis of a single element (132; 22). 32) in relation to which the extent of the geometric structure is determined.
    [6]
    Method according to claim 5, characterized in that the geometric structure (43) comprises a closed geometric figure with a certain general shape which according to the expansion rule expands by growing while maintaining the certain general shape, the geometric extent of the structure being limited in the dimensions of the data set. by allowing the shape to expand only until it reaches elements whose element properties along it. its periphery meets the said criterion.
    [7]
    A method according to claim 5, characterized (23; 33) in that the structure comprises a collection of radial parts, each of which covers a certain angular range in relation to said individual elements (23; 33). 22; 32), and in that each radial part according to the expansion rule can expand independently of the other radial parts, in that the expansion of each radial part is limited by allowing its periphery to expand only to elements whose element properties meet said criterion. k: raX7en (l2; 22; 32)
    [8]
    8. A method according to any one of 5-7, characterized in that the structure according to the expansion rule expands by, at any given stage of expansion, only sections of the periphery of the structure whose expansion is not currently limited by elements meeting said criterion are allowed to expand.
    [9]
    9. A method according to any one of claims 1-4, characterized in that the structure (54; 64) comprises an elongate part, the extent of which in a longitudinal direction is limited by not allowing it to cross elements whose element properties meet said criterion.
    [10]
    10. lO. Method according to claim 9, characterized in that the extent of the elongate part is allowed to describe a curved curve through the image.
    [11]
    11. ll. A method according to any one of the preceding claims, characterized in that a first, initial classification of the data set is performed before step b) is performed, in which initial classification elements are associated with classes in a first set of classes, and in that one or more element properties for each element are made to represent these associations so that the possible association of each element to one or more classes * in the first, set of classes can be read from the data set.
    [12]
    12. l2. Method according to claim ll, characterized in that the first classification is made to be of an unsupervised type.
    [13]
    13. A method according to any one of the preceding claims, characterized in that: the image comprises a two-dimensional photographic image, and in that each element comprises image information in at least one channel each indicating light intensity for a certain wavelength spectrum.
    [14]
    Method according to one of the preceding claims, characterized in that the geometric measurement or dimensions are made to be based on maximum radius, length, circumference, degree. of convexity, central moment, and / or area, surface area, volume, center, center of gravity the degree of circularity or sphericity of the specific geometric structure.
    [15]
    A method according to any one of the preceding claims, characterized in that, in step b), the variable geometric structure is limited in that it is not allowed to include elements which are binary associated with the property in question.
    [16]
    l6. Method according to any one of the preceding claims, characterized in that, in step kn, at least a part of the variable geometric structure. is limited by the fact that the part is not allowed to comprise a quantity comprising at least one element which together meet a predetermined criterion regarding the association of the element in question with the property in question. Application text doc 2010-0940 100172EN
    类似技术:
    公开号 | 公开日 | 专利标题
    Wang et al.2014|Mitosis detection in breast cancer pathology images by combining handcrafted and convolutional neural network features
    Schneider et al.2015|Joint 3-D vessel segmentation and centerline extraction using oblique Hough forests with steerable filters
    Zhang et al.2013|SVM-based classification of segmented airborne LiDAR point clouds in urban areas
    Narkhede2013|Review of image segmentation techniques
    US9710888B2|2017-07-18|Method and system for image analysis and interpretation
    JP4660546B2|2011-03-30|Method for characterizing objects in digitized images and computer-readable program storage
    JP2004536367A|2004-12-02|Computer-based image analysis
    Wang et al.2008|Improved classifier for computer‐aided polyp detection in CT Colonography by nonlinear dimensionality reduction
    US20180082153A1|2018-03-22|Systems and methods for deconvolutional network based classification of cellular images and videos
    US20180082104A1|2018-03-22|Classification of cellular images and videos
    Wang et al.2010|Segmentation of multispectral high-resolution satellite imagery based on integrated feature distributions
    US8781216B2|2014-07-15|Method for performing automatic classification of image information
    US20200410691A1|2020-12-31|Segmentation system for segmenting an object in an image
    Yang et al.2014|Extracting buildings from airborne laser scanning point clouds using a marked point process
    Wagner et al.2020|Deep learning and adaptive graph-based growing contours for agricultural field extraction
    Sharma et al.2016|Cell nuclei attributed relational graphs for efficient representation and classification of gastric cancer in digital histopathology
    SE1050937A1|2012-03-11|Procedure for automatically classifying a two- or high-dimensional image
    Younis et al.2019|Semantic segmentation on small datasets of satellite images using convolutional neural networks
    JP2016146132A|2016-08-12|Shape feature extraction method, shape feature extraction processing device, shape description method and shape classification method
    Florindo et al.2015|Texture descriptors by a fractal analysis of three-dimensional local coarseness
    Mylonas et al.2015|A region-based genesis segmentation algorithm for the classification of remotely sensed images
    CN107492101B|2020-06-05|Multi-modal nasopharyngeal tumor segmentation algorithm based on self-adaptive constructed optimal graph
    Staal et al.2004|Automatic rib segmentation in CT data
    Jesna et al.2013|A survey on MR Brain image segmentation using SOM based strategies
    LEE et al.2015|Grain‐oriented segmentation of images of porous structures using ray casting and curvature energy minimization
    同族专利:
    公开号 | 公开日
    SE535070C2|2012-04-03|
    US9036924B2|2015-05-19|
    US20130230257A1|2013-09-05|
    EP2614467A4|2017-12-06|
    EP2614467A1|2013-07-17|
    WO2012033460A1|2012-03-15|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US6263088B1|1997-06-19|2001-07-17|Ncr Corporation|System and method for tracking movement of objects in a scene|
    GB0103965D0|2001-02-17|2001-04-04|Univ Nottingham|Image and image content processing,representation and analysis for image matching,indexing or retrieval and database management|
    JP2008528949A|2004-11-23|2008-07-31|ケアストリームヘルスインコーポレイテッド|Automatic shape classification method|
    US7840071B2|2006-12-12|2010-11-23|Seiko Epson Corporation|Method and apparatus for identifying regions of different content in an image|
    CN101226190B|2007-01-17|2013-07-03|深圳迈瑞生物医疗电子股份有限公司|Automatic sorting method and apparatus for flow type cell art|
    WO2008118977A1|2007-03-26|2008-10-02|Desert Research Institute|Data analysis process|
    SE534089C2|2009-03-10|2011-04-26|Choros Cognition Ab|Procedure for classifying image information|WO2016071725A1|2014-11-05|2016-05-12|Umm Al-Qura University|Grading and monitoring of a geographical region|
    CN109344908B|2018-10-30|2020-04-28|北京字节跳动网络技术有限公司|Method and apparatus for generating a model|
    法律状态:
    优先权:
    申请号 | 申请日 | 专利标题
    SE1050937A|SE535070C2|2010-09-10|2010-09-10|Procedure for automatically classifying a two- or high-dimensional image|SE1050937A| SE535070C2|2010-09-10|2010-09-10|Procedure for automatically classifying a two- or high-dimensional image|
    US13/821,606| US9036924B2|2010-09-10|2011-09-09|Method for automatically classifying a two-or higher-dimensional image|
    PCT/SE2011/051095| WO2012033460A1|2010-09-10|2011-09-09|Method for automatically classifying a two- or higher-dimensional image|
    EP11823858.3A| EP2614467A4|2010-09-10|2011-09-09|Method for automatically classifying a two- or higher-dimensional image|
    [返回顶部]