Weak texture image automatic discrimination method and device based on image information ball measurement

Abstract

The invention provides a weak texture image automatic discrimination method and device based on image information ball measurement. The method comprises the steps of firstly calculating features such as a texture gradient mean value, a texture energy field radius and a texture information entropy, then analyzing image textures from three angles of a gray level average change degree, a local gray level jitter subtle degree and a global gray level confusion degree, and constructing an image texture intensity vector field to realize image texture richness description. Then constructing an information ball classifier, taking a small amount of weak texture image data obtained by visual interpretation of an expert as a prior standard of weak texture discrimination, finally constructing an information ball for data to be detected in a vector space, and calculating the probability of occurrence of an event of the data to be detected in the prior weak texture data. And generating three description components in the TRV of the to-be-measured image into the information amount of the to-be-measured data so as to realize the discrimination of the weak texture image.

Description

Technical Field

[0001] The invention belongs to a remote sensing image processing method, and specifically relates to a method and a device for automatically distinguishing weak texture images by extracting image texture richness and constructing image information spheres. Background Art

[0002] Given the richness and diversity of geographical scenes, aerial images usually acquired face complex scene information. The collected images are often mixed with scenes such as houses, cities, rural areas, waters, snowfields, etc. These images can be roughly divided into regular texture images and weak texture images. At present, a large number of studies are devoted to experiments on regular images with clear geometric structures. The analysis and application of regular texture objects such as buildings, roads, and vegetation have been relatively mature. However, when processing weak texture scenes in images, such as lakes, deserts, snowfields and other weak texture scenarios, the existing technology still faces a series of difficulties. Images in weak texture scenes have some unique characteristics, including low reflectivity, high neighborhood similarity, and insignificant differences between regions. In the actual production process, since the two types of images have significant texture and geometric differences, it is often necessary to distinguish and process the two types of images. Therefore, in order to serve the full intelligent automation of the production process, it is necessary to classify and preprocess the initial collected images, and how to quantitatively extract the texture information of the image and construct a texture classifier is the key, so it is necessary to study this.

[0003] Image texture discrimination and classification is a research hotspot in the field of computer vision, and plays an important role in target detection, image segmentation and other fields. Most of the existing methods for image texture classification can be divided into two types: traditional algorithms and deep learning algorithms. Traditional texture discrimination and classification methods are mainly divided into four categories: based on filters, statistical features, structural features and model features. Filter methods such as Gabor filters can effectively extract image features, but they are highly dependent on parameters; statistical methods such as gray-level co-occurrence matrices are only suitable for small-scale images; model methods such as fractal geometry perform well in processing complex textures, but are sensitive to image transformations. In recent years, some improved algorithms such as ILQP and NF-LBP, as well as multi-color texture descriptors, have proposed solutions to the limitations of these traditional methods, but it is still difficult to effectively process weak texture images. Texture classification methods based on deep learning have stronger feature abstraction capabilities than traditional methods, which improves classification accuracy. Autoencoders and convolutional neural networks (CNNs) were the focus of early research, and CapsNets and dual attention mechanism networks have further improved feature extraction and classification performance. In recent years, researchers have proposed a variety of innovative methods, such as dynamic texture analysis combined with RNN models, fuzzy network models based on gray-level co-occurrence matrix and particle swarm optimization (FuzzyNet), etc. These methods have performed well in specific applications. However, deep learning methods still face challenges such as data requirements and label accuracy in practical applications.

[0004] In summary, a large number of studies have been conducted on image texture discrimination and classification. However, both traditional methods and deep learning methods focus on conventional texture images and aim to discriminate certain concrete objects, and cannot effectively discriminate weak textures in images. Based on this, the present invention proposes a method for quantifying image texture richness and constructing an image information ball classifier to achieve automatic discrimination of weak texture images. Summary of the invention

[0005] The present invention proposes a method and device for automatically distinguishing weak-texture images based on image information spherical measurement, which are used to solve the problem of distinguishing weak-texture remote sensing images.

[0006] The technical solution adopted by the present invention is: a method for automatically distinguishing weak texture images based on image information spherical measurement, comprising the following steps:

[0007] Step 1: Initialize the input image, complete the parameter and discrimination threshold initialization, and obtain the prior weak texture image data through visual interpretation;

[0008] Step 2: construct the texture intensity vector TRV by calculating the texture gradient mean, texture energy field radius, and texture contour information entropy of the image, and extract the quantitative result of the image texture strength;

[0009] Step 3: Input the prior weak texture image data obtained by visual interpretation, construct the respective TRV vectors, and obtain the prior knowledge of weak texture semantics;

[0010] Step 4: Generate a vector space based on the TRV vector in step 3, and construct an information sphere in the vector space for the data to be measured. The radius of the information sphere is calculated by the minimum circumscribed sphere of the prior weak texture image data.

[0011] Step five, construct an information ball classifier, calculate the probability of the test data in the information ball under weak texture semantics, and calculate the amount of information of the test data under weak texture semantics according to Shannon's theorem, and finally compare it with the set threshold to realize the texture strength judgment of the test data.

[0012] Furthermore, in step 2, the image texture gradient mean is calculated based on the Scharr filter result to reflect the average change degree of the image grayscale, where the Scharr filter formula is shown in formula (1):

[0013]

[0014] In formula (1), G x and G t Represents the gradient of the pixel in the x and y directions, and P1…9 represents the intensity of adjacent pixels in the 3×3 neighborhood of the image;

[0015] Calculate the mean value of image texture gradient and evaluate the degree of overall image gradient change. The relevant calculation formula is shown in formula (2):

[0016]

[0017] In formula (2), represents the mean value of the texture gradient of the image, w and h represent the width and length of the image gradient map, G i Represents the gradient intensity value of pixel i, G x and G y Represents the gradient of the pixel in the x and y directions.

[0018] Furthermore, in step 2, the Laplace filter is first used to sharpen the edge features of the image, filter out noise and image information with smooth changes; then, the frequency domain conversion is performed under the fast Fourier transform to obtain the spectrum of the image; then, by analyzing the spectrum, the intensity value range and the upper quartile intensity value are calculated; then, the extreme distance point farthest from the center point is found according to the upper quartile intensity value; finally, the texture energy field is constructed with the extreme distance point as the boundary to obtain the energy field radius;

[0019] The fast Fourier transform calculation is shown in formula (3):

[0020]

[0021] In formula (3), f is the function representation of the signal in the original spatial domain, F is the function representation of the signal in the frequency domain after transformation, u and x are both from 0 to K-1, W 2K is a sine and cosine function, K is a positive integer, and F(u) is a Fourier transform, which is expressed as the sum of the products of a finite number of coefficients and sine and cosine functions;

[0022] Based on the spectrum graph, the upper quartile intensity value of the spectrum graph is used as the threshold to calculate the high-frequency radiation range of the image. The upper quartile intensity value and the texture energy field radius are calculated as shown in formulas (4) and (5):

[0023] Q up =(max(I)-min(I))·75% (4)

[0024] In formula (4), Q up is the upper quartile intensity value in the spectrum graph, max(I) and min(I) are the operations of obtaining the maximum and minimum intensity values ​​in image I;

[0025]

[0026] In formula (5), S pnt For all intensity values ​​in the spectrum graph, Qup A collection of point coordinates, Int i is the spectrum intensity value of pixel i, S dis For S pnt The distance from the midpoint to the center of the spectrum. x0 and y0 are the coordinates of the center of the spectrum. i and i is the coordinate of pixel i, R text is the texture energy field radius of the image.

[0027] Furthermore, in step 2, the image contour information entropy is calculated based on the Canny filtered image contour map to reflect the degree of global grayscale disorder. The image information entropy is calculated as shown in formula (6):

[0028]

[0029] In formula (6), H is the image information entropy, P i is the probability of occurrence of gray value i in the image, CNT i is the number of occurrences of pixels at gray level i, h and w are the height and width of the image.

[0030] Furthermore, in step 4, the vector space uses the linearly normalized Euclidean distance for position measurement, and the calculation is shown in formula (7):

[0031]

[0032] In formula (7), S dis is the similarity distance set of images, where N is the number of prior weak texture images, and D j represents the normalized Euclidean similarity distance between the test data and the prior weak texture image j, n is the vector dimension, D uv is the normalized Euclidean similarity distance between images u and v, x u 、x v is the TRV of images u and v, x u i is the i-th component value of TRV of image u; x u i ' is the normalized TRV value of the i-th component of image u, x v i ' is the i-th component value of the normalized TRV of image v, min(x i ) and max(x i ) is the minimum and maximum value of the i-th component in all TRV data.

[0033] Furthermore, the calculation formula of the information sphere radius in step 4 is shown in (8):

[0034]

[0035] In formula (8), Ten th is the space tension threshold, i.e. the information sphere radius, is the tension threshold Ten th The weak texture set under g n is a priori weak texture image in the weak texture set in the vector space, is the similarity distance set of the nth image.

[0036] Furthermore, in step 5, the probability of the data to be tested in the information ball occurring under weak texture semantics is calculated as shown in formula (9):

[0037]

[0038] In formula (9), S weak represents a weak texture set, P(x|S weak ) is the test data x in the weak texture set S weak The conditional probability under R (x) is the number of prior weak texture images contained in the information sphere of the data to be tested, and max represents the maximum value.

[0039] Furthermore, in step 5, the information content of the data to be tested is calculated by the occurrence probability of the data to be tested, and the calculation formula is shown in (10):

[0040] I(x|S weak )=-log2(P(x|S weak )) (10)

[0041] In formula (10), I(x|S weak ) is the amount of information in the data to be tested, P(x|S weak ) is the test data x in the weak texture set S weak The conditional probability under ;

[0042] Finally, the shadow information ball classifier is constructed according to the information amount of the data to be tested, as shown in formula (11):

[0043]

[0044] In formula (11), B text is the final texture discrimination result of the image, 1 means it is a weak texture; I(x|S weak ) is the amount of information of the data to be tested; σ th is the information threshold.

[0045] Furthermore, it also includes step six, using confusion matrix, accuracy, precision and recall rate to verify and analyze the discrimination results.

[0046] The present invention also provides an automatic identification device for weak texture images based on spherical measurement of image information, including a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the automatic identification method for weak texture images based on spherical measurement of image information as described in the above scheme.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] The weak texture remote sensing image discrimination method proposed in the present invention is divided into two parts: image texture richness vector extraction and information ball classifier construction. First, the texture of the image is analyzed from three perspectives: the average grayscale change, the subtlety of local grayscale jitter, and the global grayscale chaos, through the texture gradient mean, texture energy field radius, and texture information entropy. A texture richness vector is constructed to analyze and describe the texture of the image. Then, an information ball classifier is constructed, and the sparse weak texture image data obtained by visual interpretation is used as the evaluation criterion for weak texture discrimination. Then, an information ball is constructed for the data to be tested in the vector space. By calculating the probability of occurrence of events in the data to be tested in the prior weak texture data, the three descriptive components in the TRV of the image to be tested are condensed into the information amount of the data to be tested, and finally the weak texture image is discriminated. The results show that the method proposed in the present invention can better realize the discrimination of weak texture remote sensing images and is more stable than traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 Flowchart for calculating the radius of the texture energy field;

[0051] Figure 3 It is a visualization diagram of the texture energy field results of the data to be tested;

[0052] Figure 4 Construct a result graph for the test data TRV;

[0053] Figure 5 Construct a schematic diagram for the image information sphere of the data to be measured;

[0054] Figure 6 Schematic diagram of a weak texture to be tested and a non-weak texture to be tested;

[0055] Figure 7 Schematic diagram of the quantitative confusion matrix. DETAILED DESCRIPTION

[0056] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0057] Please see Figure 1 Flowchart, a method for distinguishing weak texture remote sensing images provided by the present invention comprises the following steps:

[0058] Step 1: Initialize the input remote sensing image, complete the parameter and discrimination threshold initialization, and calibrate a small number of weak texture image samples through expert visual interpretation to obtain prior weak texture image data.

[0059] Step 2: Analyze the texture of the image by calculating the texture richness vector TRV (texture richness vector) of the image, and extract the quantitative result of the image texture strength.

[0060] Preferably, the texture richness of the image in step 1 is quantitatively described. The TRV of the present invention is formed by three texture statistical information components: texture gradient mean, texture energy field radius, and texture profile information entropy. The image is texture analyzed from the average grayscale change, the subtlety of local grayscale jitter, and the degree of global grayscale confusion. The texture distribution of the image is characterized by the condensed TRV, and the texture richness vector is constructed quickly, concisely, and thoroughly, providing a basic guarantee for the construction of the texture classifier below.

[0061] Step 3: Calculate the mean of image gradient. The image gradient map is calculated by Scharr filtering, and then the image gradient mean is obtained by image mean calculation. In the visual nervous system, image texture is intuitively manifested as the difference in intensity values ​​between image pixels and surrounding pixels. When the image pixels and adjacent pixels change significantly, the image is understood to have a strong texture. In computer vision, in order to highlight the change in pixel values ​​between adjacent pixels, a gradient map is often constructed for the image, and the gradient value is used to characterize the rate of change and direction of the pixel value in the image. Therefore, by constructing a gradient map of the image and characterizing the change in pixel values ​​between adjacent pixels of the image, the image texture information can be effectively highlighted. Furthermore, the image gradient map is condensed into a mean indicator to analyze the image texture from a quantitative perspective.

[0062] Therefore, the present invention calculates the gradient map of the image by Scharr filtering. The definition of Scharr filtering is shown in formula (1):

[0063]

[0064] In formula (1), G x and G yRepresents the gradient of the pixel in the x and y directions. P1…9 represents the intensity of adjacent pixels in the 3×3 neighborhood of the image.

[0065] Calculate the mean value of the image texture gradient and evaluate the degree of change of the overall image gradient. The relevant calculation formula is shown in formula (2).

[0066]

[0067] In formula (2), represents the mean value of the texture gradient of the image, w and h represent the width and length of the image gradient map, G i Represents the gradient intensity value of pixel i, G x and G y Represents the gradient of the pixel in the x and y directions.

[0068] Step 4: Calculate the radius of the image texture energy field. First, use the Laplace filter to sharpen the edge features of the image, filter out noise and image information with smooth changes; then perform frequency domain conversion under the fast Fourier transform to obtain the spectrum of the image; then calculate the intensity value range and the upper quartile intensity value by analyzing the spectrum; then find the extreme distance point farthest from the center point based on the upper quartile intensity value; finally, construct the texture energy field with the extreme distance point as the boundary to obtain the energy field radius. Figure 2 As shown in Figure 1, a schematic diagram of texture energy field calculation using remote sensing images is shown. The fast Fourier transform calculation is shown in formula (3).

[0069]

[0070] In formula (3), f is the function representation of the signal in the original spatial domain, F is the function representation of the signal in the frequency domain after transformation (the image can be regarded as a signal), u and x are both 0 to K-1, W 2K is a sine and cosine function, K is a positive integer, and F(u) is a Fourier transform, which can be expressed as the sum of the products of a finite number of coefficients and sine and cosine functions.

[0071] Based on the spectrum, the upper quartile intensity value of the spectrum is used as the threshold to calculate the high-frequency radiation range of the image. The upper quartile intensity value and the texture energy field radius are calculated as shown in formulas (4) and (5).

[0072] Q up =(max(I)-min(I))·75% (4)

[0073] In formula (4), Q up is the upper quartile intensity value in the spectrum graph, and max(I) and min(I) are operations for obtaining the maximum and minimum intensity values ​​in image I.

[0074]

[0075] In formula (5), S pnt For all intensity values ​​in the spectrum graph, Q up A collection of point coordinates, Int i is the spectrum intensity value of pixel i, S dis For S pnt The distance from the midpoint to the center of the spectrum. x0 and y0 are the coordinates of the center of the spectrum. i and i is the coordinate of pixel i, R text is the texture energy field radius of the image. The texture energy field radius calculated by the above steps is as follows Figure 3 As shown, Figure 3 The calculation results of 10 regular images and 10 weak texture images are shown in Figure 2.

[0076] Step 5: Calculate the image texture information entropy. The image contour information entropy is calculated based on the image contour map of the Canny filter. The Canny filter extracts image contour features through Gaussian filtering, gradient calculation, non-maximum gradient suppression, and dual threshold connection analysis, and can accurately extract image edge contours. Information entropy is a statistical feature description of the source and a measure of the degree of chaos of events in the system. As a digital signal, the image can use information entropy to describe the chaos of the grayscale distribution in the image, reflecting the richness of the color in the image. The calculation of image information entropy is shown in formula (6).

[0077]

[0078] In formula (6), H is the image information entropy, P i is the probability of occurrence of gray value i in the image, CNT i is the number of occurrences of pixels at gray level i, h and w are the height and width of the image. So far, the texture richness quantification of the image is completed. The constructed TRV vector is as follows: Figure 4 shown.

[0079] Step 6: So far, the texture richness of the image is quantified, and the abstract image texture information is reduced to a vector representation. Then, the prior weak texture image data obtained by visual interpretation is input, and the TRV vectors of these data are constructed to obtain the prior knowledge of weak texture semantics. Through the quantitative definition of weak texture, it is used as the basis for subsequent image classification and judgment.

[0080] Step 7: Construct an image information sphere. Input N prior weak texture images and unknown data to be tested. In the present invention, the value of N is 8. By constructing the TRV of the prior weak texture image data, the image texture information is expressed as a spatial point in the vector space. Then, an information sphere in the space is constructed for the data to be tested. The radius of the sphere is calculated by the minimum circumscribed sphere of the prior weak texture image data. The position calculation in the vector space adopts the normalized Euclidean distance, and the calculation formula is as follows:

[0081]

[0082] In formula (7), S dis is the similarity distance set of images, where N is the number of prior weak texture images, and D j represents the normalized Euclidean similarity distance between the image to be tested and the prior weak texture image j; n is the vector dimension. In the present invention, since the TRV vector contains three components, the value of n is 3, and D uv is the normalized Euclidean similarity distance between images u and v, x u 、x v is the TRV of images u and v, x u i is the i-th component value of TRV of image u. u i ' is the normalized TRV value of the i-th component of image u, x v i ' is the i-th component value of the normalized TRV of image v, min(x i ) and max(x i ) is the minimum and maximum value of the i-th component in all TRV data.

[0083] The similarity distance set shows the spatial semantic gap between a certain image and the rest of the images. The smaller the distance, the closer the semantics. Then, the maximum semantic distance in the input prior weak texture data is calculated and used as the sphere radius of the subsequent test data to ensure that at this distance, the semantic information of the N input prior weak texture data in the vector space remains consistent. The calculation formula is shown in (8).

[0084]

[0085] In formula (8), Ten th is the space tension threshold, i.e. the information sphere radius, is the tension threshold Ten th The weak texture set under is the similarity distance set of the nth image, g nIt is a priori weak texture image in the weak texture set in the vector space. Ten refers to the tension when constructing the information sphere in the space, which is the radius of the sphere in mathematical sense. Figure 5 The data information sphere to be tested constructed through the above steps is shown.

[0086] Step 8: Construct an information ball classifier. The semantics of the data to be tested is analyzed by the number of weak texture sample points contained in the information ball to be tested. First, the occurrence probability of the data to be tested under weak texture semantics is calculated through the information ball. The image occurrence probability calculation is shown in formula (9).

[0087]

[0088] In formula (9), S weak represents a weak texture set, P(x|S weak ) is the data to be tested x in S weak The conditional probability under R (x) is the number of prior weak texture images contained in the information sphere of the data to be tested, is the maximum m existing in the prior weak texture image data R (x): Since the present invention inputs 8 pieces of prior weak texture data, the value is 8.

[0089] Then, according to Shannon's theorem, the information content of the data to be tested is calculated by the probability of occurrence. The calculation formula is shown in (10).

[0090] I(x|S weak )=-log2(P(x|S weak )) (10)

[0091] In formula (10), I(x|S weak ) is the amount of information in the data to be tested, P(x|S weak ) is the weak texture set S of the prior data in the data to be tested x weak The conditional probability of .

[0092] Finally, the information ball classifier is constructed according to the information amount of the data to be tested, as shown in formula (11).

[0093]

[0094] In formula (11), B text is the final texture discrimination result of the image, 1 means it is a weak texture; I(x|S weak ) is the amount of information of the data to be tested; σ th is the information threshold, which is a non-fixed value. In the present invention, the value is taken as 0.5 for experimental verification.

[0095] Step 9: Use confusion matrix, accuracy, precision and recall rate to verify and analyze the discrimination results of the present invention. The present invention uses 80 remote sensing images to test the performance of the algorithm, including 40 regular texture images and 40 weak texture images. Some data sets are shown in Figure 6 The proposed weak texture remote sensing image discrimination method is named MBISC, and compared with several optimal image classification and discrimination methods (Ori-KMC, Ori-HC, LBP-KMC and LBP-HC). The comparison results are shown in Table 1.

[0096] Table 1 Index analysis table

[0097]

[0098] Figure 7 The confusion matrix results of 80 images under different discrimination methods are shown. All methods can accurately discriminate non-weak texture images, but there are differences in the discrimination of weak texture images. The Ori-KMC and Ori-HC methods have a low success rate in discriminating weak textures, mainly due to interference from noise and environmental factors. The LBP-KMC and LBP-HC algorithms improve the discrimination success rate through filtering processing, but there are still about 1 / 4 misjudgments. Table 1 shows that although various methods have high accuracy in the classification of non-weak texture images, the problem of missed detection of weak textures is still prominent. In contrast, the MBISC method of the present invention can accurately discriminate 40 conventional non-weak textures and 40 weak texture images, and all three evaluation indicators reach 100%, with the best performance, proving its effectiveness in discriminating weak texture images.

[0099] On the other hand, an embodiment of the present invention also provides an automatic identification device for weak texture images based on spherical measurement of image information, including a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the automatic identification method for weak texture images based on spherical measurement of image information as described in the above scheme.

[0100] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0101] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A method for automatically distinguishing weak texture images based on image information sphere measurement, characterized in that: The following steps are involved: Step 1: Initialize the input image, complete the parameter and discrimination threshold initialization, and obtain the prior weak texture image data through visual interpretation; Step 2: construct the texture intensity vector TRV by calculating the texture gradient mean, texture energy field radius, and texture contour information entropy of the image, and extract the quantitative result of the image texture strength; Step 3: Input the prior weak texture image data obtained by visual interpretation, construct the respective TRV vectors, and obtain the prior knowledge of weak texture semantics; Step 4: Generate a vector space based on the TRV vector in step 3, and construct an information sphere in the vector space for the data to be measured. The radius of the information sphere is calculated by the minimum circumscribed sphere of the prior weak texture image data. Step five, construct an information ball classifier, calculate the probability of the test data in the information ball under weak texture semantics, and calculate the amount of information of the test data under weak texture semantics according to Shannon's theorem, and finally compare it with the set threshold to realize the texture strength judgment of the test data.

2. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: In step 2, the image texture gradient mean is calculated based on the Scharr filter result to reflect the average change degree of the image grayscale. The Scharr filter formula is shown in formula (1): In formula (1), G x and G y Represents the gradient of the pixel in the x and y directions, and P1…P9 represents the intensity of adjacent pixels in the 3×3 neighborhood of the image; Calculate the mean value of image texture gradient and evaluate the degree of overall image gradient change. The relevant calculation formula is shown in formula (2): In formula (2), represents the mean value of the texture gradient of the image, w and h represent the width and length of the image gradient map, G i Represents the gradient intensity value of pixel i, G x and G y Represents the gradient of the pixel in the x and y directions.

3. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: In step 2, the Laplace filter is first used to sharpen the edge features of the image, filter out noise and image information with smooth changes; then the frequency domain is converted under the fast Fourier transform to obtain the spectrum of the image; then the intensity value range and the upper quartile intensity value are calculated by analyzing the spectrum; then the extreme distance point farthest from the center point is found according to the upper quartile intensity value; finally, the texture energy field is constructed with the extreme distance point as the boundary to obtain the energy field radius; The fast Fourier transform calculation is shown in formula (3): In formula (3), f is the function representation of the signal in the original spatial domain, F is the function representation of the signal in the frequency domain after transformation, u and x are both from 0 to K-1, W 2K is a sine and cosine function, K is a positive integer, and F(u) is a Fourier transform, which is expressed as the sum of the products of a finite number of coefficients and sine and cosine functions; Based on the spectrum graph, the upper quartile intensity value of the spectrum graph is used as the threshold to calculate the high-frequency radiation range of the image. The upper quartile intensity value and the texture energy field radius are calculated as shown in formulas (4) and (5): Q up =(max(I)-min(I))·75% (4) In formula (4), Q up is the upper quartile intensity value in the spectrum graph, max(I) and min(I) are the operations of obtaining the maximum and minimum intensity values ​​in image I; In formula (5), S pnt For all intensity values ​​in the spectrum graph, Q up A collection of point coordinates, Int i is the spectrum intensity value of pixel i, S dis For S pnt The distance from the midpoint to the center of the spectrum. x0 and y0 are the coordinates of the center of the spectrum. i and i is the coordinate of pixel i, R text is the texture energy field radius of the image.

4. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: In step 2, the image contour information entropy is calculated based on the Canny filtered image contour map to reflect the degree of global grayscale disorder. The image information entropy is calculated as shown in formula (6): In formula (6), H is the image information entropy, P i is the probability of occurrence of gray value i in the image, CNT i is the number of occurrences of pixels at gray level i, h and w are the height and width of the image.

5. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: In step 4, the vector space uses the linearly normalized Euclidean distance to measure the position, and the calculation is shown in formula (7): In formula (7), S dis is the similarity distance set of images, where N is the number of prior weak texture images, and D j represents the normalized Euclidean similarity distance between the test data and the prior weak texture image j, n is the vector dimension, D uv is the normalized Euclidean similarity distance between images u and v, x u 、x v is the TRV of images u and v, x u i is the i-th component value of TRV of image u; x u i ' is the normalized TRV value of the i-th component of image u, x v i ' is the i-th component value of the normalized TRV of image v, min(x i ) and max(x i ) is the minimum and maximum value of the i-th component in all TRV data.

6. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 5, characterized in that: The calculation formula of the information sphere radius in step 4 is shown in (8): In formula (8), Ten th is the space tension threshold, i.e. the information sphere radius, is the tension threshold Ten th The weak texture set under g n is a priori weak texture image in the weak texture set in the vector space, is the similarity distance set of the nth image.

7. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: In step 5, the probability of the data to be tested in the information ball occurring under weak texture semantics is calculated as shown in formula (9): In formula (9), S weak represents a weak texture set, P(x|S weak ) is the test data x in the weak texture set S weak The conditional probability under R (x) is the number of prior weak texture images contained in the information sphere of the data to be tested, and max represents the maximum value.

8. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: In step 5, the information content of the data to be tested is calculated by the probability of occurrence. The calculation formula is shown in (10): I(x|S weak )=-log2(P(x|S weak )) (10) In formula (10), I(x|S weak ) is the amount of information in the data to be tested, P(x|S weak ) is the test data x in the weak texture set S weak The conditional probability under ; Finally, the shadow information ball classifier is constructed according to the information amount of the data to be tested, as shown in formula (11): In formula (11), B text is the final texture discrimination result of the image, 1 means it is a weak texture; I(x|S weak ) is the amount of information of the data to be tested; σ th is the information threshold.

9. The method for automatically distinguishing weak texture images based on image information spherical measurement according to claim 1, characterized in that: It also includes step six, using confusion matrix, accuracy, precision and recall rate to verify and analyze the discrimination results.

10. An automatic identification device for weak texture images based on image information spherical measurement, characterized in that: The invention comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the method for automatically distinguishing weak texture images based on image information spherical measurement as described in any one of claims 1 to 9.

Images