Image texture analysis method and device

Abstract

The present disclosure relates to a technique for analyzing texture of an image, and provides an image texture analysis apparatus including: an image preprocessor that extracts a region of interest to be analyzed from an image, and performs boundary enhancement on the region of interest to calculate a preprocessed image; a textural feature extractor that calculates corresponding textural feature values of the pre-processed image using different algorithms, and generates a histogram for the pre-processed image based on the textural feature values; and a texture quantizer that calculates an entropy of the histogram to calculate a texture value of the image.

Description

TECHNICAL FIELD

[0001] The disclosure relates to a technology for analyzing texture of an image. BACKGROUND

[0002] Image analysis means extracting meaningful information from an image. Primarily, a specific object, feature, situation, etc. is extracted from an image by a digital image processing technology, and the image is analyzed on this basis.

[0003] In particular, as the importance of image analysis increases in various businesses, such as autonomous driving technology using a camera device, and the development of artificial intelligence technology using an image, various technologies for analyzing an image have been developed.

[0004] In addition, as secondary battery technology advances, a technology of capturing the shape of a material (particle) in an image by using a device such as a material microscope to study the physical properties of the material (particle) has also been developed.

[0005] For object recognition, object classification, etc. in an image, computer processing provides results with less difference between users. However, texture analysis of an object in an image has no specific standard and is highly sensitive to various factors, and thus still has a large difference between users. SUMMARY

[0006] Technical problem

[0007] An object of exemplary embodiments is to provide a technology for quantifying texture features to quantify an object in an image.

[0008] TECHNICAL SOLUTION

[0009] An exemplary embodiment of the disclosure derived to solve the above object can provide an image texture analysis device including an image pre-processor that extracts a region of interest to be analyzed from an image and performs boundary enhancement on the region of interest to calculate a pre-processed image.

[0010] In addition, exemplary embodiments can provide an image texture analysis device including a texture feature extractor that calculates texture feature values of a pre-processed image using different algorithms and generates a histogram for the pre-processed image based on the texture feature values.

[0011] In addition, exemplary embodiments can provide an image texture analysis device including a texture quantifier that calculates a texture numerical value of an image by calculating an entropy of a histogram.

[0012] Further, the exemplary embodiments can provide a method for analyzing image texture, which includes extracting a region of interest to be analyzed from an image, and performing a boundary enhancement on the region of interest to calculate a pre-processed image.

[0013] In addition, the exemplary embodiments can provide a method for analyzing image texture, which includes calculating respective texture feature values of a pre-processed image using different algorithms, and generating a histogram for the pre-processed image based on the texture feature values.

[0014] In addition, the exemplary embodiments can provide a method for analyzing image texture, which includes calculating a texture numerical value of an image by calculating an entropy of a histogram.

[0015] Advantages

[0016] According to the exemplary embodiments, quantized texture feature values for objects in an image can be digitized. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 FIG. 1 is a schematic diagram for describing a configuration of an image texture analysis apparatus according to exemplary embodiments.

[0018] Figure 2 FIG. 3 is a schematic diagram for describing an operation of extracting a region of interest according to exemplary embodiments.

[0019] Figure 3 FIG. 5 is a schematic diagram for describing a boundary enhancement operation according to exemplary embodiments.

[0020] Figures 4a to 4c FIG. 7 is a schematic diagram showing each result of an image pre-processing operation.

[0021] Figure 5 FIG. 9 is a schematic diagram for describing a texture feature extraction operation according to exemplary embodiments.

[0022] Figure 6a FIG. 11 is a schematic diagram showing a local binary pattern map of an image according to exemplary embodiments.

[0023] Figure 6b FIG. 13 is a schematic diagram showing a fractal dimension map of an image according to exemplary embodiments.

[0024] Figure 7 FIG. 15 is a schematic diagram for describing a quantization result for each image according to exemplary embodiments.

[0025] Figure 8 FIG. 17 is a schematic diagram for describing an image texture analysis method according to exemplary embodiments. DETAILED DESCRIPTION

[0026] Hereinafter, some exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. When a component is added with a reference numeral to each drawing, the same components can have the same reference numeral as possible even though the same components are illustrated in different drawings. Also, in the description of the exemplary embodiments, detailed descriptions of related known configurations or functions incorporated herein can be omitted when it is determined that the subject matter of the present disclosure can be ambiguous. The terms such as "include", "have", and "consist of" used herein are generally intended to allow addition of other components, unless the terms are used with the term "only". When a component is expressed in singular form, the singular form can include the plural form unless it is explicitly stated otherwise.

[0027] When describing components of the exemplary embodiments of the present disclosure, terms including first, second, A, B, (a), (b), and the like can be used. These terms are merely intended to distinguish a component from other components, and the terms do not limit the nature, order or sequence of the component.

[0028] When describing a positional relationship of components, when two or more components are described as being "connected", "coupled", or "joined", it should be understood that the two or more components can be directly "connected", "coupled", or "joined", but can also be "connected", "coupled", or "joined" through further "insertion" of other components in addition to the two or more components. Here, the other components can be included in one or more components of the two or more components that are "connected", "coupled", or "joined" to each other.

[0029] When describing a time flow relationship related to components, methods of operation, or methods of manufacturing, for example, when the time or flow relationship is described as "after", "after both", "next", "before", or the like, unless "immediately" or "directly" is used, a discontinuous case can also be included.

[0030] Meanwhile, when referring to a numerical value or corresponding information of a component (for example, a level, etc.), the numerical value or corresponding information can be interpreted to include an error range that can occur due to various factors (for example, a processing factor, an internal or external influence, noise, etc.) even if there is no separate explicit description.

[0031] Hereinafter, exemplary embodiments will be described with reference to the accompanying drawings.

[0032] When an image is acquired through an apparatus capable of photographing a fine particle, for example, a material microscope, various property features can be extracted from the corresponding image. Also, features of the image can be used to infer features of the fine particle.

[0033] However, unlike the image of the general macro environment, it is difficult to clearly specify the boundary in the image photographed in the micro environment, and it can be difficult to objectively perform the feature between images due to the shape of the irregular particle cluster, etc.

[0034] In particular, in the case of various analysis devices for determining the physical properties of particles, it is difficult to objectify the physical properties of particles by mutual comparison due to differences between analysis devices, differences between users who identify the shape and distribution of particles, etc.

[0035] In this regard, the present disclosure aims to provide an image analysis technique that realizes objective comparative analysis by converting and providing the features of an image as numerical values. In particular, a method for solving the objectification difficulty due to differences between users and differences between analysis devices by quantifying and providing the texture features of an image is provided.

[0036] Although the present disclosure is described based on an image, it can be equally applied to image texture analysis photographed by an image. For example, the texture of an image can be quantified by converting a photographed image into an image based on a specific frame and applying the techniques described below. In addition, the following will be described based on an image of a particle photographed, but is not limited thereto. That is, even when various objects are photographed, the techniques according to the present disclosure can be applied.

[0037] Figure 1 is a schematic diagram for explaining a configuration of an image texture analysis device according to an exemplary embodiment.

[0038] Referring to Figure 1 , the image texture analysis device 100 can include an image pre-processor 110 that extracts a region of interest to be analyzed from an image and performs boundary enhancement on the region of interest to calculate a pre-processed image.

[0039] For image analysis, a pre-processing procedure can be required on a received image. The pre-processing procedure can include various operations.

[0040] For example, the image pre-processor 110 can perform an operation for extracting a region of interest to be analyzed. The region of interest refers to a region configured so that only an object to be analyzed is specified among objects and a background included in an image.

[0041] The image pre-processor 110 can binarize the image based on the calculated segmentation boundary value to calculate a binary image, and extract a region of interest by morphological operation to correct the binary image.

[0042] The image can include a background as well as one or more objects (e.g., particles). In order to accurately extract a region of interest, it is important to clearly define a boundary between an object and a background.

[0043] To this end, the image pre-processor 110 calculates a binary image and finely corrects the calculated binary image to extract a region of interest.

[0044] For example, the image pre-processor 110 can calculate a segmentation boundary value to distinguish a region of interest (e.g., an object) from a region of non-interest (e.g., a background). The segmentation boundary value can be calculated by applying an algorithm for extracting various boundary values. For example, an otsu algorithm can be used to divide pixels into two groups by randomly defining a boundary value, repeatedly calculate a contrast distribution of the two groups, and then calculate a boundary value that makes the contrast distribution of the two groups most uniform as a segmentation boundary value. In addition, various algorithms used can also be applied to calculate a boundary of an image.

[0045] Meanwhile, the image pre-processor 110 can binarize the image based on the calculated segmentation boundary value to calculate a binary image.

[0046] In addition, the image pre-processor 110 can extract a region of interest by correcting a binary image through a morphological operation. The morphological operation can be applied to more finely extract a region of interest from a binary image calculated based on a segmentation boundary value. The applied morphological operation can sequentially include a closing operation and an opening operation. For example, by performing an erosion operation after a dilation operation through a closing operation, empty spaces in the image are filled, and noise can be removed from the image in which the empty spaces have been filled by performing a dilation operation after an erosion operation through an opening operation. Thereby, a more detailed region of interest can be derived.

[0047] Meanwhile, in order to objectify a texture analysis result of an image, a bias according to a setting of a device for capturing the image needs to be considered. According to the setting of the device, even for the same material, signal intensity in the image can not be shown with the same signal. Therefore, when the texture of the image is quantified and provided, it is necessary to adjust the image before image analysis to enable comparison between values.

[0048] For example, the image pre-processor 110 can divide the image in the region of interest into blocks having a predetermined size and convert the image into a gray scale image by performing histogram equalization on each block. For example, the image pre-processor 110 can equalize the image using a contrast limited adaptive histogram equalization (CLAHE) algorithm.

[0049] For the converted gray scale image, a boundary enhancement operation can be performed by the image pre-processor 110. The texture analysis performance of the image in the region of interest can be improved by boundary enhancement.

[0050] For example, the image pre-processor 110 can perform a boundary enhancement by applying a Canny edge algorithm to the image in the region of interest, and normalize the boundary-enhanced image to a predetermined signal intensity range to calculate a pre-processed image. Specifically, the image pre-processor 110 can remove noise from the image by applying a Gaussian filter. Further, the image pre-processor 110 filters the denoised image in a horizontal direction and a vertical direction by a Sobel kernel. The magnitude and direction of a gradient are calculated from the two filtered images. Then, the image pre-processor 110 performs a non-maximum suppression operation of selecting only pixels having a local maximum gradient size as edge pixels to eliminate a phenomenon that one edge is represented by a plurality of pixels. Finally, the image pre-processor 110 selects a final edge based on a minimum threshold and a maximum threshold through a hysteresis edge tracking operation.

[0051] The image pre-processor 110 can normalize the boundary-enhanced image to pre-process a normalized image such that a signal intensity is shown in a predetermined range.

[0052] Accordingly, the image pre-processor 110 can extract a region of interest in detail from the acquired image, enhance a boundary, and then normalize the image to calculate a pre-processed image for extracting an image feature.

[0053] The image texture analysis apparatus 100 can include a texture feature extractor 120 that calculates respective texture feature values of the pre-processed image using different algorithms, and generates a histogram of the pre-processed image based on the texture feature values.

[0054] The texture feature extractor 120 calculates respective texture feature values of the image using different algorithms, and generates a histogram of the pre-processed image using the respective calculated texture feature values. Conventional algorithms for extracting texture features are sensitive to a scale such that objective quantification is difficult, or are sensitive to a change in a gray scale such that objective quantification is difficult. That is, in the case of using a specific algorithm, an image texture analysis result is sensitive to a specific factor (e.g., a scale, etc.) such that it is impossible to objectify texture features by comparing with each other.

[0055] To solve this problem, in an exemplary embodiment, texture feature values are calculated using different algorithms, and one histogram is generated using all of the respective calculated texture feature values to objectify texture features of an image.

[0056] For example, the different algorithms can be a local binary pattern algorithm and a fractal dimension algorithm.

[0057] The texture feature extractor 120 can calculate the texture feature value for each image coordinate by applying a local binary pattern algorithm to the preprocessed image. Also, the texture feature extractor 120 can generate a fractal dimension map by applying a difference box counting method to the preprocessed image, thereby calculating a signal intensity in the fractal dimension map for each image coordinate as the texture feature value.

[0058] Also, the texture feature extractor 120 can identify one or more coordinates in the preprocessed image having the same texture feature value calculated by the local binary pattern algorithm, and generate a histogram using the texture feature values calculated by the fractal dimension algorithm corresponding to the one or more coordinates.

[0059] For example, the texture feature extractor 120 can generate a histogram by summing the texture feature values calculated by the fractal dimension algorithm corresponding to the one or more coordinates having the same texture feature value calculated by the local binary pattern algorithm, and mapping the summed texture feature value to a histogram bin content on the same texture feature value calculated by the local binary pattern algorithm.

[0060] Thereby, a histogram including both the rotation, gray scale invariant feature by the local binary pattern algorithm and the scale invariant feature by the fractal dimension map can be calculated. Accordingly, it is possible to analyze the texture feature capable of objectivizing the mutual comparison by preventing a change in texture quantization due to a rotation or a scale difference, etc.

[0061] Meanwhile, the image texture analysis apparatus 100 can include a texture quantifier 130 that calculates an entropy of the histogram to calculate a texture value of the image.

[0062] For example, the texture quantifier 130 can calculate an entropy value based on the histograms generated using different algorithms to quantify and output the texture feature of the image. For example, the texture quantifier 130 can calculate the entropy of the image using a bin probability in the histogram. Also, various well-known algorithms can be used as the algorithm for calculating the entropy of the histogram.

[0063] Through this operation, the image texture analysis apparatus 100 can provide the texture of the object in the region of interest of the image as a value. Also, the image texture analysis apparatus 100 processes the image to have a feature robust to apparatus settings, image scales, rotations, etc., and provides the image texture as a value based on this. Accordingly, it is possible to compare the textures between images according to the objectivized image texture features, predict physical properties, etc., based on the image texture features.

[0064] Hereinafter, the operation of the image texture analysis apparatus described above will be described in more detail with respect to each step with reference to the accompanying drawings.

[0065] Figure 2 is a schematic diagram for describing an operation of extracting a region of interest according to an exemplary embodiment.

[0066] Referring to Figure 2 When an image to be analyzed is received from an image generation device, the image texture analysis device can extract a region of interest to be analyzed from a corresponding image.

[0067] The image pre-processor can binarize the image based on the calculated segmentation boundary value to calculate a binary image, and extract the region of interest by correcting the binary image via a morphological operation.

[0068] The image pre-processor calculates a segmentation boundary value using an otsu algorithm based on the received image (S200).

[0069] For example, the received image can include a background, a plurality of objects, etc. The image pre-processor needs to extract a region of interest to separate an object to be analyzed from the background. To this end, the image pre-processor calculates a segmentation boundary value that serves as a segmentation criterion.

[0070] For example, the segmentation boundary value can be calculated using an otsu algorithm. The otsu algorithm refers to an algorithm for finding an appropriate threshold when thresholding a certain image. The image pre-processor randomly defines a boundary value of the received image to divide pixels into two groups, repeatedly calculates a contrast distribution of the two groups, and then selects a boundary value that makes the contrast distribution of the two groups most uniform. That is, if a certain threshold is T, the image pre-processor can calculate the segmentation boundary value by calculating the best T that has the smallest difference in the ratio of pixels that are binarized based on T.

[0071] The image pre-processor calculates a binary image based on the calculated segmentation boundary value (S210). When the segmentation boundary value is calculated, the image pre-processor converts and calculates the received image into a binary image based on the calculated segmentation boundary value. The converted binary image can be distinguished by binarizing an object region and a background region in the image. However, in order to set a more clear and more detailed region of interest, it is necessary to correct the region of interest in the binary image.

[0072] To this end, the image pre-processor refines the binary image by a morphological operation to extract a region of interest (S220). The morphological operation method refers to a method for analyzing and processing the shape of an image. Thereby, it is possible to clearly indicate an unclear region of interest that occurs during binarization processing of the image such as hole filling and noise removal.

[0073] Specifically, if the original image and the binary image are and , the region of interest can be obtained as follows .

[0074] [Formula 1]

[0075]

[0076] Here, the OPENING operation, the CLOSING operation, and the DOT operation correspond to morphological opening, morphological closing, and inner product in morphological operations, respectively.

[0077] The morphological operations include an erosion operation and a dilation operation. The open operation is an algorithm that removes protrusions from a binary image and corrects narrow breaks to smooth a contour after the erosion operation using the dilation operation. The close operation is an algorithm that fills in recessed portions or small holes in a binary image to smooth a contour and maintain the shape and size of an object after the dilation operation using the erosion operation.

[0078] Accordingly, the image pre-processor performs the close operation to smoothly correct a boundary in a binary image, and performs the open operation on the corrected image to remove protruding noise and correct breaks, thereby clearly correcting a contour.

[0079] Through these operations, a region of interest to be analyzed for image texture in a received image can be more clearly derived.

[0080] Meanwhile, in the case of a microparticle, a boundary of an object for texture analysis needs to be enhanced for the inside of the particle. In particular, a microparticle can have various forms of texture, and due to limitations in size, image resolution, etc., a boundary of a shape of an object for texture analysis can be blurred. Accordingly, for high-precision texture analysis, boundary enhancement of an object in a region of interest is required.

[0081] Figure 3 is a schematic diagram for describing a boundary enhancement operation according to an exemplary embodiment.

[0082] Referring to Figure 3 , the image pre-processor performs an adaptive histogram equalization operation to smooth a region of interest image to a gray level (S300).

[0083] In a video or an image, depending on the settings of a device, even for the same material, signal intensity is rarely presented in the same signal. Accordingly, the image pre-processor performs histogram equalization on a region of interest to process signal intensity of the same material to have the same gray value.

[0084] In the case of a conventional histogram equalization algorithm, the distribution of all pixels in an image is considered and uniformly applied. In this case, a brightness difference within an image is not considered, which causes a problem in which shape recognition of an important object becomes more difficult.

[0085] In consideration of these problems, the image pre-processor can divide the image in the region of interest into blocks having a predetermined size and convert the image into a gray scale image by performing histogram equalization on each block. In addition, when the histogram equalization is completed for each block, the boundary between blocks can be smoothed by applying bilinear interpolation. As an example, the image pre-processor processes signal intensities of the same material to have the same gray scale value through a histogram height limit (clip limit) of contrast limited adaptive histogram equalization (CLAHE).

[0086] The image pre-processor performs a noise removal operation (S310). The image pre-processor removes noise by applying a Gaussian filter. For example, if there is noise in the image, it can be difficult to properly find an edge. Therefore, before the boundary enhancement process, a N x N Gaussian filter is used to remove noise in the image. Here, N can also be set to 1 or a natural number greater than or equal to 1.

[0087] The image pre-processor performs a boundary enhancement operation (S320). For example, the image pre-processor can perform boundary enhancement by applying a Canny edge algorithm to the image within the region of interest.

[0088] The image pre-processor obtains a gradient in each direction by applying a Sobel kernel to the image from which noise is removed using a Gaussian filter in the horizontal direction and the vertical direction. The gradient at the (x, y) coordinates of the image can be obtained by Equation 2 below.

[0089] [Equation 2]

[0090]

[0091]

[0092] Here, the horizontal gradient is , and the vertical gradient is .

[0093] After obtaining the gradient, the image pre-processor scans the entire image to remove pixels that do not contribute to the edge. In scanning the image, the image pre-processor finds a pixel having the largest gradient value in the scanning region in the gradient direction. In this case, in order to eliminate the phenomenon in which one edge is represented by a plurality of pixels, only a pixel having a local maximum gradient size is selected as an edge pixel.

[0094] Finally, the image pre-processor determines whether the selected edge is actually an edge. For this, a minimum threshold value and a maximum threshold value are required. If the gradient intensity is greater than the maximum value, the selected edge becomes an edge, and if the gradient intensity is less than the minimum value, the selected edge is discarded. By confirming connectivity with other edges, a value between the minimum boundary and the maximum boundary is selected as a final edge.

[0095] The boundary enhancement operation has been described above based on the Canny edge algorithm. However, this is for ease of understanding, and various boundary enhancement algorithms can be applied to the exemplary embodiments. That is, there is no limitation on any algorithm capable of enhancing the boundary of the image in the region of interest.

[0096] Further, the image pre-processor can normalize the boundary enhanced image to a predetermined range of signal intensity to calculate the pre-processed image.

[0097] For example, the image pre-processor can perform an operation of adding the boundary enhanced image and normalizing it to the original image.

[0098] If the region of interest image is referred to as , the boundary enhanced image can be acquired as Equation 3 .

[0099] [Equation 3]

[0100]

[0101] Here, is a detailed parameter indicating the degree of enhancement of the boundary image and can be set in advance. is the result of the Canny edge algorithm. NORMALIZE refers to a process of normalizing the signal intensity to the range of 0 to 255.

[0102] Through this operation, the image pre-processor can enhance and normalize the boundary of the region of interest of the image to acquire a grayscale image in a predetermined signal intensity range.

[0103] Figures 4a to 4c is a schematic diagram showing each result of the image pre-processing operation.

[0104] Figure 4a An original image acquired from an image acquisition device in the image texture analysis device is shown. As Figure 4a shown, the original image can include both a background region and a particle. Here, since texture analysis is based on the particle, the region of interest is set through the image pre-processing process.

[0105] Figure 4b A binary image of the region of interest configuring the region of interest described above is shown. As described above, the region of interest is refined through binary imaging and morphological operations.

[0106] Figure 4c A pre-processed image having an enhanced and normalized boundary for the region of interest image (Ipre) is shown. ). As shown in FIG. 4C, when the pre-processing procedure according to an exemplary embodiment is performed, a background portion other than the particulate is excluded from the region of interest, and a normalized gray scale image having an enhanced boundary of a particulate portion to be analyzed for texture analysis is obtained.

[0107] The image texture analysis apparatus extracts texture feature values based on the pre-processed image.

[0108] Figure 5 is a schematic diagram for describing a texture feature extraction operation according to an exemplary embodiment.

[0109] Referring to Figure 5 , the texture feature extractor receives a pre-processed image (S500). Here, the pre-processed image means a gray scale image having an enhanced and normalized boundary as described above.

[0110] The texture feature extractor can calculate respective texture feature values using different algorithms (S510 and S520).

[0111] For example, the texture feature extractor can extract texture feature values through a local binary pattern (LBP) algorithm (S510). The local binary pattern compares signal intensities between a center pixel and N neighboring pixels within a certain region. Based on the comparison result, the value of each pixel is binarized, listed, and converted into a decimal value to calculate a pixel value. When N is 8, the decimal value is calculated as 8 binary digits, and the value of each coordinate is calculated as a value between 0 and 255. In an exemplary embodiment, a rotation-invariant LBP can be applied. Alternatively, various LBP algorithms, for example, a square LBP, a uniform LBP, a circular LBP, and a uniform rotation-invariant LBP, can also be applied.

[0112] Specifically, for example, for a pre-processed image, when a pixel position (x, y) and a signal size are given, the texture feature extractor samples N pixels within a radius R based on (x, y), and performs a procedure of calculating a local binary pattern value using the sampled pixels, as shown in Equation 4.

[0113] [Equation 4]

[0114]

[0115] Here, a sign function (sign) is calculated as in Equation 5.

[0116] [Equation 5]

[0117]

[0118] The local binary pattern value can have a value between 0 and .

[0119] The texture feature extractor calculates local binary pattern values of the preprocessed image as the texture feature values according to the local binary pattern algorithm described above.

[0120] Figure 6a is a schematic diagram illustrating a local binary pattern map of an image according to an exemplary embodiment.

[0121] Referring to Figure 6a The local binary pattern map can be acquired by applying the local binary pattern algorithm to the preprocessed image.

[0122] However, when a histogram is generated using only the local binary pattern values, there is a disadvantage that the histogram is sensitive to scale changes. Therefore, in the exemplary embodiment, a histogram is generated by further using a fractal dimension algorithm having a characteristic that is invariant to scale changes.

[0123] To this end, the texture feature extractor can extract texture feature values through the fractal dimension algorithm (S520).

[0124] For example, the texture feature extractor can perform the process of generating a Gaussian scale space and the process of calculating a fractal dimension map by performing a difference box count (DBC).

[0125] For example, the texture feature extractor applies a plurality of Gaussian scale parameters to the preprocessed image to generate a plurality of Gaussian scale spaces. The Gaussian scale parameter is a parameter that determines the degree of blurring (the degree of Gaussian scale) of the image, and has a plurality of values according to settings. Therefore, a plurality of Gaussian scale spaces are generated according to the number of Gaussian scale parameters.

[0126] Thereafter, the texture feature extractor performs an operation of calculating a difference between a maximum value and a minimum value within a region in each Gaussian scale space and calculating a gradient within each Gaussian scale space.

[0127] For example, if the degree of Gaussian scale according to the Gaussian scale parameter is r, the plurality of Gaussian scale spaces can be expressed as G(x, y, r).

[0128] The reason for calculating the Gaussian scale space is that a change in r can be confirmed at each (x, y) position. A gradient of a linear regression of the r value at each (x, y) position is used as a value of F(x, y).

[0129] [Equation 6]

[0130]

[0131] Referring to Equation 6, G(x, y, r) denotes a value in the Gaussian scale space, F(x, y) denotes an intensity (gradient) of the fractal dimension, r denotes a degree of blurring, which is a degree of Gaussian scale, and b denotes a regression constant.

[0132] Through this operation, the texture feature extractor can calculate the intensity G(x, y, r) using the Gaussian scale (r) applied at each coordinate and the intensity F(x, y) in the fractal dimension map.

[0133] The signal intensity at the (x, y) coordinate in the fractal dimension map is calculated as F(x, y), which is set as a texture feature value calculated by a fractal dimension algorithm.

[0134] Figure 6b FIG. 7 is a schematic diagram illustrating a fractal dimension map of an image according to an exemplary embodiment.

[0135] Referring to Figure 6b The fractal dimension map can be acquired by applying a fractal dimension algorithm to the preprocessed image.

[0136] However, the texture feature value calculated by the fractal dimension algorithm is robust to scale variation, but has a problem of being sensitive to gray scale variation. Accordingly, in the exemplary embodiment, a histogram is generated by additionally using a local binary pattern texture feature value having a characteristic robust to gray scale variation.

[0137] The texture feature extractor can generate a histogram of the preprocessed image based on the texture feature value (S530).

[0138] As an example, the texture feature extractor can identify one or more coordinates in the preprocessed image having the same texture feature value calculated by the local binary pattern algorithm, and generate a histogram using the texture feature value calculated by the fractal dimension algorithm corresponding to the one or more coordinates.

[0139] For example, the texture feature extractor can generate the histogram by summing the texture feature value calculated by the fractal dimension algorithm corresponding to the one or more coordinates, and mapping the texture feature value to a histogram bin content on the same texture feature value calculated by the local binary pattern algorithm.

[0140] That is, the texture feature extractor can correct a histogram of the local binary pattern by applying the texture feature value calculated by the fractal dimension algorithm to the histogram of the local binary pattern.

[0141] [Equation 7]

[0142]

[0143] In Equation 7, denotes a corrected histogram of the local binary pattern. denotes an intensity at a (x, y) position of the fractal dimension map. p means a texture feature value denoting a local binary pattern result.

[0144] For example, when a local binary pattern algorithm is executed, a texture feature value p at each coordinate has a value from 0 to The texture feature extractor finds the coordinates of (x, y) having the same p value in the local binary pattern map. For example, if the coordinates of (x, y) of a local binary pattern texture feature value p = 1 are { (0, 0), (2, 5)}, the texture feature extractor generates a value of the sum of the texture feature values at the (x, y) positions of the fractal dimension map corresponding to the coordinates. That is, the texture feature extractor generates an F (0, 0) + F (2.5) value, and maps the corresponding value to a histogram bin on p = 1 to generate a histogram.

[0145] Accordingly, the texture feature extractor finds the texture feature values in the fractal dimension map while changing the p value from 0 to and sums them to generate a histogram.

[0146] Through these operations, the texture feature extractor can generate a histogram having scale-invariant, gray-scale-invariant, and rotation-invariant features. This provides an effect of making the texture of an image digital and objective, and enabling a relative comparison.

[0147] Meanwhile, the image texture analysis apparatus can further include a texture quantifier for quantifying and expressing the texture. The texture quantifier calculates an entropy of the image using the bin probabilities in the histogram generated above, and quantifies the calculated entropy.

[0148] Various algorithms can be used to perform the operation of calculating the entropy from the histogram.

[0149] For example, the texture quantifier can calculate the entropy using a histogram generated using texture feature values calculated by different algorithms.

[0150] In the case of a continuous random variable X having a probability density function f(x), the definition of entropy is shown in Equation 8.

[0151] [Equation 8]

[0152]

[0153] Here, H(X) is the entropy, and f(x) is the probability density function. In order to apply the definition of the entropy calculation to the case where the density is expressed as a histogram, if the range of the variable is divided into n intervals, , k = 1,..., n, the entropy can be calculated as shown in Equation 9 below.

[0154] [Equation 9]

[0155]

[0156] Then, the k-th term of the sum of Equation 9 is associated with the k-th bin of a histogram having a width In the histogram, the bin probability , k = 1,..., n can be defined as shown in Equation 10.

[0157] [Equation 10]

[0158]

[0159] Thus, the bin probability in the histogram can be approximated by a rectangular area having a height Here, is a representative value within the interval Similarly, the k-th term of the sum of Equation 9 can be approximated by

[0160] In other words, in terms of the bin probability, the entropy can be calculated as shown in Equation 11.

[0161] [Equation 11]

[0162]

[0163] In the case where is constant but need not be equal to 1, the entropy can be expressed as shown in Equation 12 below.

[0164] [Equation 12]

[0165]

[0166] When is not constant, adjustment is required on a bin-by-bin basis to compare histograms having different bin configurations in order to apply Equation 11. The continuous random variable can be calculated as a natural logarithm.

[0167] Through this operation, the texture quantizer can use the bin probabilities in the histogram to calculate the entropy of the image.

[0168] In addition, the texture quantizer can calculate the entropy by using an entropy calculation algorithm for various known histograms to perform quantization of the texture of the image.

[0169] Figure 7 is a graph for describing numerical results of each image according to an exemplary embodiment.

[0170] Referring to Figure 7 , results of quantifying texture according to an exemplary embodiment are shown for various acquired images.

[0171] For example, Figure 7 ​700 has an entropy value of 2.3313094545036425. 710 is expressed as 2.4270400444038986, and 720 is expressed as 2.564448503174378. 730 has the highest entropy value of 2.7513417037248868.

[0172] Accordingly, the entropy value varies according to the texture of the target image, and thus an objective comparative analysis of the textures of images with each other can be achieved. In particular, in the case of material particles, it is possible to quantify the correlation between the texture of a material and physical characteristics, which can provide various applications, such as predicting a physical property through a texture, predicting a texture through a physical property, etc.

[0173] Hereinafter, the operation of the image texture analysis apparatus described above will be briefly described again. The operation of the image texture analysis apparatus described above can be performed in each step to be described below. In addition, for ease of understanding, each step to be described below is divided, and integration or division of each step also falls within the scope of the exemplary embodiments.

[0174] Figure 8 is a schematic diagram for describing an image texture analysis method according to an exemplary embodiment.

[0175] Referring to Figure 8 , the method for analyzing an image texture can include extracting a region of interest to be analyzed from an image, and performing boundary enhancement on the region of interest to calculate a preprocessed image (S800).

[0176] For example, in S800, an operation of extracting a region of interest to be analyzed can be performed. The region of interest refers to a region configured to designate only an object to be analyzed among objects and a background included in an image.

[0177] Extracting the region of interest can be a step of binarizing an image based on a calculated segmentation boundary value to calculate a binary image, and correcting the binary image through a morphological operation to extract the region of interest. The image can include a background and one or more objects (for example, particles). In order to accurately extract the region of interest, it is important to clearly define the boundary between the object and the background. To this end, extracting the region of interest is a step of calculating a binary image and correcting the calculated binary image in detail to extract the region of interest.

[0178] For example, the extracting the region of interest can be a step of calculating a segmentation boundary value to distinguish between a region of interest (e.g., an object) and a region of non-interest (e.g., a background). Various algorithms for extracting a boundary value can be applied to calculate the segmentation boundary value. For example, an otsu algorithm can be used to divide pixels into two groups by randomly defining a boundary value, repeatedly calculate a contrast distribution of the two groups, and then calculate a boundary value that makes the contrast distribution of the two groups most uniform as a segmentation boundary value. In addition, various algorithms to be used can also be applied to calculate a boundary from an image.

[0179] Meanwhile, the extracting the region of interest can be a step of calculating a binary image by binarizing the image based on the calculated segmentation boundary value. In addition, the extracting the region of interest can be a step of correcting the binary image by a morphological operation to extract the region of interest. The applied morphological operation can sequentially include a close operation and an open operation. For example, by the close operation, empty spaces in the image are filled by performing an erosion operation after a dilation operation, and by the open operation, noise is removed from the image in which the empty spaces have been filled by performing a dilation operation after an erosion operation. Thereby, a more detailed region of interest can be derived.

[0180] Meanwhile, depending on the settings of the device, even for the same material, the signal intensity in the image can not be shown with the same signal. Therefore, when the texture of the image is quantified and provided, it is necessary to adjust the image before image analysis so that comparison between values can be achieved.

[0181] For example, the performing the boundary enhancement can be a step of dividing the image in the region of interest into blocks having a predetermined size, and converting the image into a gray scale image by performing histogram equalization for each block. For example, the performing the boundary enhancement can be a step of equalizing the image using a contrast limited adaptive histogram equalization (CLAHE) algorithm.

[0182] For the converted gray scale image, a boundary enhancement operation can be performed by performing the boundary enhancement. The texture analysis performance of the image in the region of interest can be improved by the boundary enhancement.

[0183] For example, performing boundary enhancement can be done by applying the Canny edge algorithm to the image in the region of interest, and then normalizing the boundary-enhanced image to a predetermined signal strength range to calculate the preprocessed image. Specifically, performing boundary enhancement can be a step of removing noise from the image by applying a Gaussian filter. Additionally, performing boundary enhancement involves filtering the denoised image using Sobel kernels in both the horizontal and vertical directions. The magnitude and direction of the gradients are calculated using the two filtered images. Then, performing boundary enhancement involves performing a non-maximum suppression operation, which selects only pixels with locally maximum gradient magnitudes as edge pixels to eliminate the phenomenon of many pixels representing a single edge. Finally, performing boundary enhancement involves selecting the final edge based on minimum and maximum thresholds using a hysteresis edge tracking operation. Performing boundary enhancement can also be a step of normalizing the boundary-enhanced image to preprocess the image so that the signal strength falls within a certain range.

[0184] Therefore, extracting the region of interest and performing boundary enhancement can be a series of steps that involve extracting the region of interest in detail from the acquired image, enhancing the boundary, and then normalizing the boundary to compute a preprocessed image for extracting image features.

[0185] Meanwhile, the method for analyzing image texture may include using different algorithms to calculate the corresponding texture feature values ​​of the preprocessed image, and generating a histogram for the preprocessed image based on the texture feature values ​​(S810).

[0186] Conventional algorithms used to extract texture features are sensitive to scale, making objective quantification difficult, or sensitive to grayscale changes, making objective quantification difficult. In other words, when using a specific algorithm, the results of image texture analysis are sensitive to specific factors (such as scale), making it impossible to objectify texture features through comparison.

[0187] To address this issue, in an exemplary implementation, different algorithms are used to compute texture feature values, and all the corresponding computed texture feature values ​​are used to generate a histogram to objectify the texture features of the image.

[0188] For example, different algorithms could be local binary mode algorithms and fractal dimension algorithms.

[0189] Calculating the corresponding texture feature values ​​can be achieved by applying a local binary mode algorithm to the preprocessed image to calculate texture feature values ​​for each image coordinate. Alternatively, calculating the corresponding texture feature values ​​can be achieved by generating a fractal dimension map by applying a difference box counting method to the preprocessed image, and calculating the signal intensity in the fractal dimension map as the texture feature value for each image coordinate.

[0190] Alternatively, generating a histogram can be achieved by: identifying one or more coordinates in the preprocessed image that have the same texture feature values ​​calculated using a local binary mode algorithm, and using the texture feature values ​​corresponding to the one or more coordinates calculated using a fractal dimension algorithm to generate a histogram.

[0191] For example, generating a histogram can be achieved by the following steps: summing texture feature values ​​calculated by a fractal dimension algorithm corresponding to one or more coordinates having the same texture feature values ​​calculated by a local binary mode algorithm, and mapping the summed texture feature values ​​to the contents of a histogram bin about the same texture feature values ​​calculated by the local binary mode algorithm.

[0192] Therefore, histograms can be calculated for both rotation-invariant grayscale features obtained through the local binary pattern algorithm and scale-invariant features obtained through the fractal dimension map. Thus, texture features that allow for objective comparison can be analyzed by preventing changes in texture quantization caused by rotation or scale differences.

[0193] Methods for analyzing image texture may include calculating the texture value of an image by calculating the entropy of its histogram (S820).

[0194] For example, calculating texture values ​​can be achieved by: calculating entropy values ​​based on histograms generated using different algorithms to quantify and output the texture features of the image. Alternatively, calculating texture values ​​can be achieved by: calculating the image entropy using the bin probabilities in the histogram. Furthermore, various well-known algorithms can be used to calculate entropy from histograms.

[0195] The above operations can be used to calculate texture values ​​by providing the texture of objects in the region of interest of an image as numerical values. Alternatively, methods for analyzing image texture can include processing images with robust characteristics to device settings, image scale, rotation, etc., and then providing the image texture as numerical values ​​based on these characteristics. Therefore, it is possible to compare the textures of images based on objective image texture features, and to predict physical properties based on these features.

[0196] The above description merely illustrates the technical spirit of this disclosure, and those skilled in the art can make various changes and modifications without departing from the essential characteristics of this disclosure. Therefore, the exemplary embodiments are not intended to limit the technical spirit of this disclosure, but rather to describe it, and the scope of this disclosure is not limited by the exemplary embodiments. The scope of protection of this disclosure should be interpreted by the appended claims, and all technical concepts within their equivalents should be interpreted as falling within the scope of this disclosure.

[0197] Cross-reference to Related Applications

[0198] This patent application claims priority under 35 U.S.C. §119(a) to Korean Patent Application No. 10-2023-0074963, filed on June 12, 2023, the entire contents of which are incorporated herein by reference. In addition, this application also claims priority for countries other than the United States for the same reason, and the entire contents thereof are incorporated herein by reference.

Claims

1. An image texture analysis apparatus comprising: an image pre-processor that extracts a region of interest to be analyzed from an image, and performs boundary enhancement on the region of interest to calculate a pre-processed image; a texture feature extractor that calculates respective texture feature values of the pre-processed image using different algorithms, and generates a histogram for the pre-processed image based on the texture feature values; and a texture quantifier that calculates a texture value of the image by calculating an entropy of the histogram. The image pre-processor binarizes the image based on a calculated segmentation boundary value to calculate a binary image, and extracts the region of interest by correcting the binary image via morphological operations.

2. The image texture analysis device of claim 1, wherein, The morphological operations sequentially include a close operation and an open operation.

3. The image texture analysis device of claim 2, wherein, The image pre-processor divides the image in the region of interest into blocks having a predetermined size, and converts the image into a gray scale image by performing histogram equalization for each block.

4. The image texture analysis device of claim 1, wherein, The image pre-processor performs the boundary enhancement by applying a Canny edge algorithm to the image in the region of interest, and normalizes the boundary-enhanced image to a predetermined signal intensity range to calculate the pre-processed image.

5. The image texture analysis device of claim 1, wherein, The different algorithms are a local binary pattern algorithm and a fractal dimension algorithm.

6. The image texture analysis device of claim 1, wherein, The texture feature extractor identifies one or more coordinates in the pre-processed image having the same texture feature value calculated by the local binary pattern algorithm, and generates the histogram using the texture feature value calculated by the fractal dimension algorithm corresponding to the one or more coordinates.

7. The image texture analysis device of claim 6, wherein, The texture feature extractor generates the histogram by summing the texture feature values calculated by the fractal dimension algorithm corresponding to the one or more coordinates, and mapping the summed texture feature value to a histogram bin content with respect to the same texture feature value calculated by the local binary pattern algorithm.

8. The image texture analysis device of claim 7, wherein, The texture quantifier calculates an entropy of the image using bin probabilities in the histogram.

9. The image texture analysis device of claim 1, wherein, 10.A method for analyzing image texture, comprising: extracting a region of interest to be analyzed from an image, and performing boundary enhancement on the region of interest to calculate a pre-processed image; calculating respective texture feature values of the pre-processed image using different algorithms, and generating a histogram for the pre-processed image based on the texture feature values; and calculating a texture value of the image by calculating an entropy of the histogram. The extracting the region of interest includes binarizing the image based on a calculated segmentation boundary value to calculate a binary image, and extracting the region of interest by correcting the binary image via morphological operations.

11. The method for analyzing image texture according to claim 10, wherein, The performing the boundary enhancement includes performing the boundary enhancement by applying a Canny edge algorithm to the image in the region of interest, and normalizing the boundary-enhanced image to a predetermined signal intensity range to calculate the pre-processed image. The different algorithms are a local binary pattern algorithm and a fractal dimension algorithm.

12. The method for analyzing image texture of claim 10, wherein, ​ 13. The method for analyzing image texture of claim 10, wherein, ​ 14. The method for analyzing image texture according to claim 13, wherein, Generating the histogram includes identifying one or more coordinates in the pre-processed image having the same texture feature value computed by the local binary pattern algorithm and generating the histogram using the texture feature value computed by the fractal dimension algorithm corresponding to the one or more coordinates.

15. The method for analyzing image texture according to claim 14, wherein, Generating the histogram includes generating the histogram by summing the texture feature values computed by the fractal dimension algorithm corresponding to one or more coordinates and mapping the summed texture feature value to a histogram bin content with respect to the same texture feature value computed by the local binary pattern algorithm.

16. The method for analyzing image texture of claim 10, wherein, Computing the texture value includes using bin probabilities in the histogram to compute an entropy of the image.

Images