Polarization image texture descriptor for representing classification similarity
By introducing multi-angle polarization local and non-local binary modes in traditional LBP encoding and constructing histogram intersection representation, the problem that the traditional encoding method cannot obtain the spatial relationship of polarized images is solved, and effective description of polarized image texture and expression of classification similarity are realized.
Patent Information
- Application Number
- CN202510195344.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-21
- Publication Date
- 2025-05-23
AI Technical Summary
The traditional LBP encoding method cannot obtain various spatial relationships in the local area, resulting in less obvious differentiation of polarized image textures, and conventional texture similarity representation methods cannot adequately express the similarity of polarized image texture classification.
By obtaining orthogonal differential images, the images are encoded using multi-angle polarization local binary mode and non-local binary mode, and a spatial histogram is constructed to represent the texture classification results. The similarity of texture classification results is measured using the histogram intersection method.
An effective description of polarized image texture is realized, which can better express the classification similarity of polarized image texture, and provides new technical means for object detection and recognition.
Smart Images
Figure CN120032174A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of polarization spectrum imaging and image processing thereof, and in particular relates to a polarization image texture descriptor representing classification similarity. Background Art
[0002] Texture is a fundamental feature of the surface of an object and an important visual clue for humans to perceive the world. In the field of computer vision and pattern recognition, the extraction of texture features is of great significance. However, due to the complex texture patterns and uncontrolled imaging conditions such as roughness, smoothness and unevenness, as well as uncontrollable factors such as image rotation, illumination, scaling and perspective changes, extracting discriminative and robust texture features is a challenging task.
[0003] Light reflected from the surface of an object will produce polarization, so polarization information becomes an indispensable part of image texture representation. Different polarization angles contain different polarization information and reflect different texture information. The traditional LBP encoding method cannot obtain various spatial relationships in local areas.
[0004] Therefore, a new technical solution is urgently needed in the prior art to solve this problem. Summary of the invention
[0005] The technical problems to be solved by the present invention are: to provide a polarization image texture descriptor representing classification similarity to solve the problem that the traditional LBP encoding method cannot obtain various different spatial relationships in the local area; the problem that the polarization image texture differentiation is not obvious when using limited polarization direction images for encoding and constructing histograms; the conventional texture similarity representation method cannot fully express and describe the classification similarity of polarization image texture and other technical problems.
[0006] A polarization image texture descriptor for indicating classification similarity, characterized in that it comprises the following steps, and the following steps are performed in sequence:
[0007] Step 1: Obtain orthogonal difference image
[0008] In the range of 0° to 180°, a polarization camera is used to collect polarization images at equal intervals. The Stokes vector parameter S 0 , S 1 , S 2 Calculate the transmitted light intensity I(α) and decompose it into the natural light component I nl (α) and the total polarized light component I fpl (α), and finally, by the full polarization light component I fpl (α) Use orthogonal difference operation to obtain the orthogonal polarization component I ⊥(α), thereby obtaining N orthogonal difference images with equal angle intervals; when calculating the central pixel and its adjacent sampling points, the pixel value I of the orthogonal difference image is used n ⊥ (i, j) Calculate the grayscale average value of the n×n pixel blocks around N orthogonal difference images Realize image denoising;
[0009] Step 2: Uniform Mode Coding
[0010] When the uniformity metric U≤2, given an image Four first-order derivative images along β = 0°, 45°, 90°, 135° Respectively expressed as The second-order derivative image is defined as According to the center pixel and neighborhood pixels The relationship between is encoded by the encoding rule f(·,·), and finally the four encoded second-order derivative images are cascaded to define it as a multi-angle polarization local binary pattern (pLBP-MA)
[0011] Step 3: Non-uniform pattern coding
[0012] When the uniformity metric U>2, the extended “eriu2” is used to encode each adjacent sampling point in the non-local region, which is defined as a multi-angle polarization non-local binary pattern. Where r is the neighborhood radius and P is the number of sampling points.
[0013] Step 4: Texture similarity representation
[0014] The polarization image texture classification result is modeled by the spatial histogram K(i,β), the histogram intersection method is used to represent the similarity between histograms, and the similarity or difference between two histograms is measured by the H set and S set in the histogram intersection S.
[0015] The formula for calculating the transmitted light intensity I(α) by the Stokes vector in step 1 is as follows:
[0016]
[0017] Decompose I(α) into natural light component I nl (α) and the total polarized light component I fpl (α) The formula is as follows:
[0018] I(α)=I nl (α)+I fpl (α)
[0019]
[0020] By fpl (α) is calculated to obtain the orthogonal polarization component I ⊥ (α) The formula is as follows:
[0021]
[0022] The histogram of the polarization image texture representation result in step 4 is:
[0023] K(i,β)={K(R i )|i=1,...,O;β=0°,45°,90°,135°}
[0024] Among them, K(R i ) is from the local area R i The extracted histogram features, given a direction β, the polarization image is spatially divided into R 1 ,...,R O Represents a rectangular area.
[0025] The similarity expression between histograms is:
[0026]
[0027] The set H is used to measure the distance between polarization image texture histograms, and the set S indicates that the closer the similarity value of polarization image texture is to 1, the higher the similarity is.
[0028] Through the above design scheme, the present invention can bring the following beneficial effects:
[0029] The present invention provides a polarization image texture descriptor containing different spatial relationships, which mainly includes the encoding steps of multi-angle polarization binary patterns. The natural light component is eliminated from the calculated orthogonal difference image within the polarization direction of 0° to 180° to obtain N orthogonal difference images with equal angle intervals; so as to ensure that the obtained polarization image can contain more texture information. In the local area, the present invention obtains the first-order derivative image of the four directions of the given angle by The non-local invention uses the extended eriu2 for encoding and constructs a histogram to obtain the texture classification result. In order to obtain more spatial information, the problem of unclear texture differentiation of polarized images is solved.
[0030] Adding spatial features in the LBP encoding process can reflect more polarization information. The present invention combines the polarization feature information of images polarized in different directions, encodes the differential image through multi-angle polarization local / non-local binary patterns, and uses histogram intersection to represent the texture classification result, providing a new technical means for target detection and recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] The present invention is further described below with reference to the accompanying drawings and specific embodiments:
[0032] Figure 1 The present invention is a schematic diagram of an orthogonal difference image in a polarization image texture descriptor representing classification similarity.
[0033] Figure 2 The present invention is a schematic diagram of a local encoding process in a polarization image texture descriptor representing classification similarity.
[0034] Figure 3 A schematic diagram of the proportion of non-local coding in a polarization image texture descriptor representing classification similarity according to the present invention.
[0035] Figure 4 The present invention is a flow chart of a polarization image texture descriptor representing classification similarity. DETAILED DESCRIPTION
[0036] In order to better understand the purpose, structure and function of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.
[0037] like Figure 1 to Figure 4 As shown, the present invention provides a polarization image texture descriptor for representing classification similarity. By analyzing the relationship between the differences in orthogonal differential polarization images at different angles and the grayscale distribution of pixels in the local neighborhood of the image, the classification results are represented by combining the histogram intersection method, thereby solving the problem that simple LBP cannot represent polarization texture information from different polarization angles.
[0038] The present invention involves two parts. The first part is mainly to obtain the orthogonal polarization component I containing the most polarization information. ⊥ (α), this method is based on Figure 1 The schematic diagram of orthogonal difference at different angles shown in FIG. 1 decomposes the transmitted light intensity after the Stokes vector is solved to calculate the full polarization light component I required by the present invention. fpl (α), and finally perform orthogonal difference calculation to obtain the orthogonal polarization components.
[0039] The second part is the encoding part of the multi-angle polarization local binary pattern (pLBP-MA). The encoding process is as follows Figure 2 As shown, through the given The four first-order derivative images along the directions of 0°, 45°, 90°, and 135° are represented as The second-order derivative image is defined as for pass Figure 2 The four different permutations and combinations are encoded by the f(·,·) rule, and we get The encoding result is obtained by the same method, and the local encoding of the second-order derivative images in the other three directions is obtained. Finally, the four encoded second-order derivative images are cascaded.
[0040] Multi-angle polarization non-local binary patterns encode non-uniform patterns into different values. These non-uniform patterns correspond to complex texture structures occupying a certain proportion. Figure 3 As shown in the figure, custom r and P are used to count the sum of the texture of the vertical incident light image. It can be seen that the mode with a U value of 4-8 accounts for a large proportion. Through statistics, it can still be concluded that the adopted "eriu2" extended encoding contains richer texture features.
[0041] like Figure 4 As shown in the figure, a flowchart of a polarization image texture descriptor representing classification similarity is given. First, by using a polarization camera to collect polarization images at equal intervals, an orthogonal difference image is calculated and obtained. During this period, the image is denoised by calculating the average gray value. The multi-angle polarization local binary pattern (pLBP-MA) and multi-angle polarization non-local binary pattern are used to obtain the orthogonal difference image. Encoding is performed, and finally the constructed spatial histogram is used to measure the similarity of texture classification results using the histogram intersection method.
[0042] Specifically, a method for acquiring a polarization difference characteristic image at an optimal angle includes the following steps, and the following steps are performed in sequence:
[0043] Step S1, obtaining an orthogonal difference image.
[0044] In the range of 0 to 180°, a polarization camera is used to collect polarization images at equal intervals. The state of polarized light can be represented by Stokes vectors S0, S1, S2, and S3, that is, S = (I, Q, U, V). Since there are very few circular polarization components in nature, S3 can be ignored. The Stokes vector is:
[0045]
[0046] The transmitted light intensity I(α) is calculated by the parameters S0, S1, and S2 of the Stokes vector and decomposed into the natural light component I nl (α) and the total polarized light component I fpl (α).
[0047] The transmitted light intensity I(α) is:
[0048]
[0049] Among them, I nl (α) is the natural light component, I fpl (α) is the total polarized light component.
[0050] Full polarization component I fpl (α) is:
[0051]
[0052] By fpl (α) Use orthogonal difference operation to obtain the orthogonal polarization component I ⊥ (α) is:
[0053]
[0054] In order to reduce the impact of noise on texture description, the pixel value I of the orthogonal difference image is used to calculate the central pixel and its adjacent sampling points. n ⊥ (i,j) calculates N orthogonal difference images around Average grayscale value of pixel block for:
[0055]
[0056] in, For calculation The average value after the pixel block, I n ⊥ (i, j) is the pixel value of the orthogonal difference image.
[0057] Step S2: uniform mode encoding.
[0058] When the uniformity metric U≤2, given an image Four first-order derivative images along β = 0°, 45°, 90°, 135° Respectively expressed as:
[0059]
[0060] Define the second derivative image is defined as:
[0061]
[0062] Where f(·,·) is the center pixel according to the encoding rule and neighborhood pixels The relationship is encoded, and the encoding rules are as follows:
[0063]
[0064] Finally, the four encoded second-order derivative images are concatenated, and the multi-angle polarization local binary pattern (pLBP-MA) is defined as:
[0065]
[0066] Step S3: non-uniform mode encoding.
[0067] When the uniformity metric U>2, the U value applies the extended "eriu2" to encode each adjacent sampling point in the non-local area, and the multi-angle polarization non-local binary mode Defined as:
[0068]
[0069] Where r is the neighborhood radius and P is the number of sampling points.
[0070] Step S4: Texture similarity representation.
[0071] The polarization image texture classification result is modeled by the spatial histogram K(i,β), and the polarization image texture representation result histogram is:
[0072] K(i,β)={K(R i )|i=1,...,O;β=0°,45°,90°,135°}
[0073] Among them, K(R i ) is from the local area R i The extracted histogram features, given a direction β, the polarization image is spatially divided into R 1 ,...,R O Represents a rectangular area.
[0074] The similarity expression between histograms is:
[0075]
[0076] The set H is used to measure the distance between polarization image texture histograms, and the set S indicates that the closer the similarity value of polarization image texture is to 1, the higher the similarity is.
[0077] The histogram intersection method can be used to represent the similarity between histograms, and the similarity or difference between two histograms can be measured by the H set and S set in the histogram intersection S.
[0078] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the enlightenment of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.
Claims
1. A polarization image texture descriptor representing classification similarity, characterized by: The method comprises the following steps, and the following steps are performed in sequence: Step 1: Obtain orthogonal difference image In the range of 0° to 180°, a polarization camera is used to collect polarization images at equal intervals; the transmitted light intensity I(α) is calculated by the parameters S0, S1, and S2 of the Stokes vector and decomposed into the natural light component I nl (α) and the total polarized light component I fpl (α), and finally, by the full polarization light component I fpl (α) Use orthogonal difference operation to obtain the orthogonal polarization component I ⊥ (α), thereby obtaining N orthogonal difference images with equal angle intervals; when calculating the central pixel and its adjacent sampling points, the pixel value I of the orthogonal difference image is used n ⊥ (i,j) calculates N orthogonal difference images around Average grayscale value of pixel block Realize denoising of images; Step 2: Uniform Mode Coding When the uniformity metric U≤2, given an image Four first-order derivative images along β = 0°, 45°, 90°, 135° Respectively expressed as The second-order derivative image is defined as According to the center pixel and neighborhood pixels The relationship between is encoded by the encoding rule f(·,·), and finally the four encoded second-order derivative images are cascaded to define it as a multi-angle polarization local binary pattern (pLBP-MA) Step 3: Non-uniform pattern coding When the uniformity metric U>2, the extended "eriu2" is used to encode each adjacent sampling point in the non-local region, which is defined as a multi-angle polarization non-local binary pattern. Where r is the neighborhood radius and P is the number of sampling points. Step 4: Texture similarity representation The polarization image texture classification result is modeled by the spatial histogram K(i,β), the histogram intersection method is used to represent the similarity between histograms, and the similarity or difference between two histograms is measured by the H set and S set in the histogram intersection S.
2. The polarization image texture descriptor representing classification similarity according to claim 1, characterized in that: The formula for calculating the transmitted light intensity I(α) by the Stokes vector in step 1 is as follows: Decompose I(α) into natural light component I nl (α) and the total polarized light component I fpl (α) The formula is as follows: I(a)=I nl (a)+I fpl (a) By fpl (α) is calculated to obtain the orthogonal polarization component I ⊥ (α) The formula is as follows:
3. The polarization image texture descriptor representing classification similarity according to claim 1, characterized in that: The histogram of the polarization image texture representation result in step 4 is: K(i,β)={K(R i )|i=1,...,O;β=0°,45°,90°,135°} Among them, K(R i ) is from the local area R i The extracted histogram features, given a direction β, the polarization image is spatially divided into R1,...,R O Represents a rectangular area. The similarity expression between histograms is: The set H is used to measure the distance between polarization image texture histograms, and the set S indicates that the closer the similarity value of polarization image texture is to 1, the higher the similarity is.