Remote sensing image classification algorithm for multi-attribute feature extraction

Abstract

The invention discloses a remote sensing image classification algorithm for multi-attribute feature extraction, and belongs to the technical field of remote sensing image classification algorithms, the algorithm decomposes the same remote sensing image into a plurality of pairs of different attribute component features according to different image texture representation types, and compared with an existing single space texture feature expression mode, the algorithm has the advantages that the classification efficiency is improved; according to the method, more diverse texture attribute feature descriptions and richer image detail information can be obtained from one image, ground feature category distinguishing is more facilitated, four representative texture representation types including content, roughness, contrast and orientation are selected to generate corresponding texture attribute components, and the texture attribute feature descriptions and the image detail information can be compared with the prior art. According to the algorithm, classification experiments are carried out on hyperspectral remote sensing images and radar remote sensing images, the effectiveness of the method is proved by experimental results, it is indicated that the method is suitable for classification of various different types of remote sensing images, and good classification performance can also be obtained under the condition of small samples.

Description

Technical Field

[0001] The invention relates to a remote sensing image classification algorithm, in particular to a remote sensing image classification algorithm for multi-attribute feature extraction, and belongs to the technical field of remote sensing image classification algorithms. Background Art

[0002] Remote sensing image classification refers to the process of classifying each pixel in a remote sensing image into a specific land object category. Relevant research shows that the use of spatial information can effectively improve the accuracy of land object classification. At present, the use of spatial information is mainly divided into two types: spatial preprocessing and post-processing.

[0003] On the one hand, spatial preprocessing strategies can extract spatial features before classification. Representative methods include morphological profile method, attribute profile method, morphological component analysis, morphological neighborhood filtering, empirical mode decomposition, and wavelet filtering.

[0004] On the other hand, spatial post-processing methods usually refer to the use of spatial information to perform regularization after pixel-level classification is completed. Representative methods include clustering, watershed segmentation, relearning algorithms, graph-based classification or superpixel methods. Both spatial pre-processing and post-processing technology have received widespread attention in the field of remote sensing image classification.

[0005] Among them, the post-processing method is to perform spatial regularization on the pixel-level classification results, that is, no new features are introduced, while the spatial preprocessing method often needs to first extract some spatial features and use this new feature expression to improve the separability of the object categories. Among them, the spatial preprocessing strategy based on image decomposition has been successfully applied to different image classification tasks. In the prior art, the authors proposed an image separation method based on morphological component analysis, which uses sparse image separation technology to decompose the image into different morphological components.

[0006] In the prior art, sparsity and morphological diversity are used as an effective feature expression and applied to blind source separation. In the prior art, the image separation method based on morphological component analysis is introduced into hyperspectral image classification and has achieved good classification results. Traditional image separation technology based on morphological component analysis usually separates an image into two component features, semantic and texture. With the help of corresponding dictionaries, sparse expression and component separation of different component features are achieved.

[0007] Existing research shows that the effectiveness of morphological component analysis is mainly reflected in the following two points:

[0008] 1. The obtained component features can effectively improve the separability of ground feature categories;

[0009] 2. Compared with the original image, the separated semantic component features show a higher signal-to-noise ratio;

[0010] However, traditional morphological component analysis only decomposes images into semantic and texture parts, but ignores the fact that images often have many different types of texture feature descriptions. Therefore, a remote sensing image classification algorithm with multi-attribute feature extraction is designed to solve the above problem. Summary of the invention

[0011] The main purpose of the present invention is to provide a remote sensing image classification algorithm for multi-attribute feature extraction. The algorithm decomposes the same remote sensing image into multiple pairs of different attribute component features according to different image texture representation types. Compared with the existing single spatial texture feature expression method, the method can obtain more diverse texture attribute feature descriptions and richer image detail information from an image, which is more conducive to the classification of ground objects. The method selects four representative texture representation types: content, roughness, contrast, and orientation to generate corresponding texture attribute components. The algorithm carries out classification experiments on hyperspectral remote sensing images and radar remote sensing images. The experimental results prove the effectiveness of the method, indicating that the method is suitable for classification of various different types of remote sensing images and can also obtain good classification performance under small sample conditions.

[0012] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0013] A remote sensing image classification algorithm for multi-attribute feature extraction includes the following steps:

[0014] Step 1: Extract sub-blocks from the image as the source of dictionary learning;

[0015] Step 2: Combine deep dictionary learning on the selected sub-tile set to construct the corresponding dictionary basis for each attribute category;

[0016] Step 3: Based on sparse coding technology, realize sparse decomposition of attribute components and solve sparse expression coefficients.

[0017] Preferably, in step 2, the learning and construction of the dictionary includes content, roughness, contrast, and directionality.

[0018] Preferably, the roughness is used to characterize edge characteristics in the image;

[0019] Component separation based on roughness attribute, the image is separated into coarse component and fine component, where the coarse component is used to extract the large-scale strong texture edge features in the image, and the fine component is used to extract the small-scale weak texture edge features in the image;

[0020] For the dictionary of coarse components, bilateral filtering technique is used to generate it;

[0021] The specific formula is shown in (3):

[0022]

[0023] Among them, g(i,j) represents the output pixel at the spatial coordinate (i,j);

[0024] f(k,l) represents the pixel at its neighborhood position (k,l);

[0025] ω(i,j,k,l) ​​represents the weight coefficient;

[0026] When the bilateral filtering technique is applied to an image, the intensity value of each image pixel is calculated by the weighted average of the intensity values ​​of other pixels in its neighborhood. The weight is determined using Gaussian distribution, as shown in formula (4):

[0027]

[0028] Right now,

[0029] The weight coefficient w(i,j,k,l) ​​depends on the product of the domain kernel and the range kernel;

[0030] σ d represents the variance of the domain kernel;

[0031] σ r represents the variance of the range kernel;

[0032] Wavelet threshold filtering technology is used to extract fine components. The specific formula (5) is as follows:

[0033]

[0034] Among them, H(x,y) is the coefficient of the horizontal subband at (x,y) after wavelet threshold filtering;

[0035] h(x,y), v(x,y), and d(x,y) correspond to the coefficients of the horizontal, vertical, and diagonal subbands at (x,y) before wavelet transform, respectively;

[0036] T h , T ν and T d are the thresholds in different directions respectively.

[0037] Preferably, the contrast uses the rate of change of the pixel color brightness value in the local neighborhood window to describe the contrast attribute;

[0038] Component separation based on contrast attribute, the image is separated into two components: high contrast and low contrast, which represent the speed of the change rate of color brightness value respectively;

[0039] Anisotropic diffusion transform is used to extract contrast component features, and its basic formula (6) is as follows:

[0040]

[0041] Among them, △ is the Laplace operator;

[0042] is the gradient operator;

[0043] div is the divergence;

[0044] c(x,y,t) is the diffusion coefficient, which is used to control the diffusion rate;

[0045] The two diffusion coefficient equations proposed by Perona and Malik are expressed as follows:

[0046]

[0047] Where K is a constant used to control the sensitivity of the edge.

[0048] Preferably, the directionality is used to statistically describe the direction information of the texture in the image;

[0049] Select two orthogonal directional attributes, horizontal and vertical, and decompose the image into two attribute components in the horizontal and vertical directions by designing dictionary descriptions of two specific directions;

[0050] The threshold filtering technology based on stationary wavelet transform is used for filtering in a specific direction. The specific formula is as follows:

[0051]

[0052] Among them, ω j,k is the high frequency subband coefficient of each level of stationary wavelet transform;

[0053] T is the set wavelet threshold.

[0054] Preferably, linelikeness, regularity and roughness are also included, and a corresponding dictionary generation algorithm is designed and applied to the multi-attribute component analysis method.

[0055] Preferably, the solution in step 3 includes setting A s and A t The dictionary corresponding to the given texture attribute features, x s and x t They are the smoothing components y s and detail componentst Sparse expression coefficients under the corresponding dictionary;

[0056] For a given image y, the following optimization equation is given:

[0057] y=y s +y t +n=A s x s +A t x t +n (10);

[0058] It can consider different categories of texture attribute features and learn to construct corresponding dictionaries A according to different image attributes. s and A t , generate image components y that represent different texture attribute characteristics s and t ;

[0059] That is, expand (10) to obtain the constrained optimization solution formula (11):

[0060]

[0061] Among them, λ 1 and λ 2 is the regularization parameter;

[0062] T s =(A T s A s ) -1 A T s and T t =(A T t A t ) -1 A T t Yes A s and A t The pseudo-inverse matrix of y s =A sxs and t =A txt Export;

[0063] In the sparse coding stage, Equation (11) is solved using the variational and augmented Lagrangian algorithms. s and A t A deep sparse dictionary learning method with total variation and hard threshold regularization constraints is used to iteratively update the relevant dictionary.

[0064] Preferably, the extraction and separation of multi-attribute component features, the design of an adaptive multi-attribute dictionary, so that each attribute feature has a sparse expression under its own dictionary, but not sparse under other attribute dictionaries, thereby achieving the discrimination and separation of different attribute features under the multi-attribute joint dictionary representation. The relevant model formula is as follows:

[0065]

[0066] In formula (12), the original image S i Separated into linear combinations of different attribute component features, and α ik Then it represents the attribute component S ik In its corresponding dictionary D k The sparse representation coefficient under ;

[0067] Therefore, the sparse representation model of attribute component analysis is used to learn multi-attribute features, which is expressed in the form of formula (13):

[0068]

[0069] Where, D: = (D 1 ,D 2 ,…,D K ) represents a multi-attribute joint dictionary, Θ(D) is used to characterize the compatibility and discriminative regularity between attribute dictionaries;

[0070] Formula (13) further includes the following steps to separate each attribute feature:

[0071] The initial dictionary base is expanded by using spectral bundles or multi-endmember spectral sets, and the spectral heterogeneity is adaptively expressed by using multi-endmember spectral sets. The learning problem of multi-attribute features is converted into a class-level sparse representation reconstruction problem, and the class-level dictionary is optimized by using the image sub-block set corresponding to each attribute feature.

[0072] After obtaining the dictionary of each attribute feature, a sparse representation model of multiple attribute features is constructed under the joint sparse representation of multiple attribute dictionaries. For multiple attribute features, a common dictionary between attribute features and a personalized dictionary for each attribute feature are constructed. The semantic connection between multiple attribute features establishes an abundance multi-attribute sparse representation model.

[0073] Preferably, the deep learning model is combined with sparse representation and dictionary learning, and the deep network structure is transformed based on the aforementioned multi-attribute sparse representation model. The relevant model is shown in formula (14):

[0074]

[0075] Beneficial technical effects of the present invention:

[0076] The present invention provides a remote sensing image classification algorithm for multi-attribute feature extraction. The algorithm decomposes the same remote sensing image into multiple pairs of different attribute component features according to different image texture representation types. Compared with the existing single spatial texture feature expression method, the method can obtain more diverse texture attribute feature descriptions and richer image detail information from an image, which is more conducive to the classification of ground objects. The method selects four representative texture representation types: content, roughness, contrast, and orientation to generate corresponding texture attribute components. The algorithm carries out classification experiments on hyperspectral remote sensing images and radar remote sensing images. The experimental results prove the effectiveness of the method, indicating that the method is suitable for the classification of multiple different types of remote sensing images and can also obtain good classification performance under small sample conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A schematic diagram of a multi-attribute component analysis algorithm according to a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction of the present invention;

[0078] Figure 2 An original image diagram according to a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction of the present invention;

[0079] Figure 3 A decomposition result diagram of texture components and detail components of content attributes according to a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction of the present invention;

[0080] Figure 4 A decomposition result diagram of coarse components and fine components oriented to roughness attributes according to a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction of the present invention;

[0081] Figure 5 A decomposition result diagram of high and low contrast components of contrast attributes according to a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction of the present invention;

[0082] Figure 6 A decomposition result diagram of horizontal and vertical directional degree components of directional degree attributes according to a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction of the present invention;

[0083] Figure 7 This is a sub-pixel abundance multi-attribute deep sparse representation network model diagram of a preferred embodiment of a remote sensing image classification algorithm for multi-attribute feature extraction according to the present invention. DETAILED DESCRIPTION

[0084] In order to make the technical solution of the present invention more clear and specific to those skilled in the art, the present invention is further described in detail below in conjunction with embodiments and drawings, but the implementation manner of the present invention is not limited thereto.

[0085] like Figure 1-Figure 7 As shown, this embodiment provides a remote sensing image classification algorithm for multi-attribute feature extraction. Given an image y with N pixels, the goal of the multi-attribute component analysis algorithm is to decompose it into two attribute components: a low-frequency texture attribute component ys and a high-frequency detail attribute component yt. These separated components can be expressed as a linear combination of the original image, as shown in formula (2):

[0086] y=y s +y t +n (2)

[0087] Where n represents the approximate residual in the image decomposition process. The traditional morphological component analysis algorithm decomposes the image into semantic components and texture components, which limits the full utilization of spatial information in remote sensing images. Considering that remote sensing images contain significant spatial information such as contrast and intensity regions and edges, more adequate feature expression can be achieved through the description of multiple texture attribute features. Therefore, we propose a feature expression algorithm based on multi-attribute component analysis to decompose and attribute express the expression containing various texture attribute features, namely content component, roughness component, contrast component, and orientation component. Multi-attribute component analysis mainly includes three steps: the first step is to extract sub-blocks from the image as the source of dictionary learning; the second step is to combine deep dictionary learning on the selected sub-block set to construct the corresponding dictionary basis for each attribute category; the third step is to realize the sparse decomposition of attribute components and the solution of sparse expression coefficients based on sparse coding technology; the whole process adopts an iterative evolutionary solution process, and the relevant dictionary is iteratively updated through total variation and hard threshold regularization technology to improve the effect of attribute decomposition.

[0088] Among the three steps above, the most critical one is the second step of dictionary learning and construction. A good dictionary basis will directly affect the sparse coding solution process. Therefore, the following focuses on the generation strategies of various texture attribute components and their corresponding dictionaries:

[0089] 1) Content;

[0090] In the multi-attribute component analysis algorithm, content attributes represent traditional texture distribution and non-texture detail attribute components. The detail components based on content attributes usually characterize the anisotropic structure in the image, including various curves, edges and contours, etc.; while for the smooth components, they mainly reflect the global statistical characteristics of the image surface, that is, they characterize the spatial texture distribution information. For content attribute components, multi-scale analysis tools such as curvelet transform, biorthogonal wavelet transform, non-decimated wavelet transform, and local ridgelet transform can be used to extract them. Here, in order to extract content attribute components, local curvelet transform is used to generate a content detail component dictionary. For the smooth components of content attributes, by comparing local discrete cosine transform and Gabor transform and other technologies, a local Gabor transform is selected to generate the corresponding attribute dictionary.

[0091] 2) Roughness;

[0092] The roughness attribute is used to characterize the edge characteristics in the image. Based on the component separation of the roughness attribute, the image can be separated into a coarse component and a fine component, wherein the coarse component is used to extract the large-scale strong texture edge features in the image, while the fine component is used to extract the small-scale weak texture edge features in the image. This characteristic attribute can better reflect the edge characteristics of the local texture in the image. For the dictionary of the coarse component, a bilateral filtering technique is used to generate it. The bilateral filter is a nonlinear filter with edge preservation and noise reduction capabilities. The filter contains two components: one part is a filter based on spatial geometric distance, and the other part is a filter based on pixel difference, that is, the difference between the central pixel and its neighboring pixel values ​​is used for characterization. The combination of these two parts enables the bilateral filter to have better edge preservation and smoothing denoising characteristics, so that the pixel value of the filter output comes from the weighted average of the neighboring pixel values. The specific formula is shown in (3):

[0093]

[0094] Where g(i,j) represents the output pixel at the spatial coordinate (i,j), f(k,l) represents the pixel at its neighborhood position (k,l), and ω(i,j,k,l) ​​represents the weight coefficient. When the bilateral filter is applied to an image, the intensity value of each image pixel is calculated by the weighted average of the intensity values ​​of other pixels in its neighborhood, and the weight is determined using a Gaussian distribution, as shown in formula (4). In other words, the weight depends not only on the spatial distance between pixels, but also on the brightness value between pixels. In the image filtered by the bilateral filter, weak texture edges with similar brightness values ​​are removed, while strong texture edges are retained, thereby highlighting the strong texture edge information in the image and making the area surrounded by these strong texture edges smoother.

[0095]

[0096] That is, the weight coefficient w(i,j,k,l) ​​depends on the product of the domain kernel (the first half) and the range kernel (the second half), σ d represents the variance of the domain kernel, σ r Represents the variance of the range kernel.

[0097] In order to obtain fine components, it is necessary to suppress strong texture edges and keep weak texture edges. Therefore, a wavelet threshold filtering technology is used to complete the extraction of fine components. The specific formula is as follows:

[0098]

[0099] Among them, H(x,y) is the coefficient of the horizontal subband at (x,y) after wavelet threshold filtering, h(x,y), v(x,y) and d(x,y) correspond to the coefficients of the horizontal, vertical and diagonal subbands at (x,y) before wavelet transform. Th, Tν and Td are thresholds in different directions. This formula gives the threshold filtering in the horizontal direction, and the same method can also be used to implement the threshold filtering of the vertical and diagonal subband coefficients.

[0100] 3) Contrast

[0101] The contrast attribute is usually used to measure the variance of the grayscale distribution. Generally, the two attribute components of high contrast and low contrast represent fast and slow intensity changes respectively. We use the rate of change of pixel color brightness values ​​in the local neighborhood window to describe the contrast attribute. Based on the component separation of the contrast attribute, the image can be separated into two components, high contrast and low contrast, which represent the speed of the rate of change of the color brightness value respectively. In order to generate the corresponding dictionary, the anisotropic diffusion transform is used to extract the contrast component features. Anisotropic diffusion, also known as PM diffusion, is a widely used image filtering technique that can reduce the interference of noise while maintaining the image detail features. The basic formula is as follows:

[0102] Among them, △ is the Laplace operator, is the gradient operator, div is the divergence, and c(x, y, t) is the diffusion coefficient, which is used to control the diffusion rate. The two diffusion coefficient equations proposed by Perona and Malik are usually used. They are both functions of the image gradient and are expressed as:

[0103]

[0104] Where K is a constant used to control the sensitivity of the edge. The setting of the K value usually adopts an empirical method or a method based on image noise. On the one hand, by using anisotropic diffusion, the texture of the area with high contrast is smoothed, while the texture of the area with low contrast is maintained; on the other hand, by changing the diffusion coefficient in formula (6) from positive to negative, the opposite effect can be obtained, thereby highlighting the texture characteristics of the low contrast area.

[0105] 4) Orientation

[0106] The directional attribute is a global attribute in the image, which is used to statistically describe the directional information of the texture in the image. Therefore, this attribute mainly reflects the overall degree of a specific texture direction in the image, and does not describe the detailed information of the specific texture direction. Considering that although the texture direction can be any direction in the range of [0,π], it can be decomposed into a combination of horizontal and vertical directions. Therefore, two orthogonal directional attributes of horizontal and vertical are selected, that is, by designing a dictionary description of two specific directions, the image is decomposed into attribute components in the horizontal and vertical directions. In order to obtain the attribute components in each direction, a threshold filtering technology based on stationary wavelet transform is used for filtering in a specific direction. The specific formula is as follows:

[0107]

[0108] Among them, ω j,k is the high-frequency subband coefficient of each level of the stationary wavelet transform, and T is the set wavelet threshold. Usually, the initial value of T is set to 0.95max(|ω|), and then it is continuously reduced with each iteration. The stationary wavelet transform overcomes the lack of translation invariance of the traditional discrete wavelet transform, and can retain more detail information in the high-frequency subband, so that even very small texture edges can be retained. By using stationary wavelet filtering to perform threshold processing on different high-frequency subband coefficients, different directional attribute components can be separated from the image.

[0109] In addition, the algorithm can further introduce other types of texture attribute features, such as linelikeness, regularity, roughness, etc., which can still be applied to our multi-attribute component analysis method by designing the corresponding dictionary generation algorithm.

[0110] In order to more clearly illustrate the proposed multi-attribute component analysis framework, a multi-attribute component analysis experiment is conducted on a 220×220 real hyperspectral image. This image is part of the hyperspectral dataset of the University of Pavia. The separation results are shown in Figure 1As shown. For each texture attribute category feature, the corresponding attribute component is generated using the above transformation. The experimental results show that compared with the traditional morphological component analysis method, the multi-attribute component analysis method proposed in this paper can bring more sufficient spatial feature information. From the experimental results, it can be observed that the contrast attribute reflects the change in density, which is very different from the content attribute. Similarly, directional features (horizontal and vertical features) provide complementary feature information. Considering that the ground objects in remote sensing images (such as roads and buildings) often have clear directionality, it is necessary to make up for the lack of expression ability of a single texture feature through more diverse texture attribute feature expressions.

[0111] Optimize the solution process;

[0112] Assume A s and A t The dictionaries corresponding to the given texture attribute features (low-frequency smooth components and high-frequency detail components), x s and x t They are the smoothing components y s and detail components t Sparse expression coefficients under the corresponding dictionary. For a given image y, the following optimization equation is given:

[0113] y=y s +y t +n=A s x s +A t x t +n (10);

[0114] The main innovation of the multi-attribute component analysis algorithm is that it can consider different categories of texture attribute features, that is, learn and construct corresponding dictionaries As and At according to different image attributes, thereby generating image components ys and yt that represent different texture attribute features. That is, expand (10) to obtain the constrained optimization solution formula (11):

[0115]

[0116] Among them, λ1 and λ2 are regularization parameters, Ts = (ATsAs)-1ATs and Tt = (ATtAt)-1ATt are pseudo-inverse matrices of As and At, which can be derived from ys = Asxs and yt = Atxt respectively. In the sparse coding stage, Equation (11) is solved by variational and augmented Lagrangian algorithms. At the same time, As and At also use the deep sparse dictionary learning method with total variation and hard threshold regularization constraints to iteratively update the relevant dictionary.

[0117] Deep sparse dictionary learning;

[0118] To achieve the extraction and separation of multi-attribute component features, the key is to design an adaptive multi-attribute dictionary so that each attribute feature has a sparse expression under its own dictionary, but not sparse under other attribute dictionaries, thereby achieving the discriminative separation of different attribute features under the multi-attribute joint dictionary representation. The relevant model formula is as follows:

[0119]

[0120] In formula (12), the original image S i Separated into linear combinations of different attribute component features, and α ik Then it represents the attribute component S ik In its corresponding dictionary D k Therefore, the sparse representation model of attribute component analysis is used to learn multi-attribute features, which can be expressed in the form of formula (13).

[0121]

[0122] Where, D: = (D 1 ,D 2 ,…,D K ) represents a multi-attribute joint dictionary, and Θ(D) is used to characterize the compatibility and discriminative regularity between attribute dictionaries. Formula (13) To separate each attribute feature, the key lies in the dictionary expression ability of various attribute features. On the one hand, each attribute dictionary is required to have the ability to represent the most significant structural characteristics of each attribute; on the other hand, the fact that multiple attributes come from the same image should not be ignored, that is, there must be a common dictionary base in the various attribute dictionaries learned. Therefore, firstly, spectral bundles or multi-endmember spectral sets are used to expand the initial dictionary base, and multi-endmember spectral sets are used to adaptively express spectral heterogeneity, so as to convert the learning problem of multi-attribute features into a class-level sparse representation reconstruction problem, and the image sub-block set corresponding to each attribute feature (that is, the set of image sub-blocks belonging to the same type of ground objects) is used to optimize the learning of the class-level dictionary; after obtaining the dictionary of each attribute feature, a sparse representation model of multi-attribute features is constructed under the joint sparse representation of multi-attribute dictionaries, and multiple attribute features are used to construct a common dictionary between attribute features and a personalized dictionary for each attribute feature, and the semantic connection between multiple attribute features is used to generate a multi-attribute sparse representation model of abundance with sparser expression, stronger discriminability and more robustness.

[0123] In general, the multi-attribute component analysis model based on sparse coding can be understood as a three-layer feature learning framework, which includes a decomposition layer for the original image and a layer for extracting various texture attributes in the image. However, this attribute feature is still a shallow feature expression, and the description of the most significant feature of the attribute by the learned dictionary is still relatively rough. Through deep dictionary learning, it can provide a natural way to describe the local details of the shallow dictionary basis learned earlier in a more refined way, which has a local amplification effect, especially when a large number of basis functions are required, it can better eliminate the overfitting phenomenon of the dictionary basis. Therefore, in order to combine the deep learning model with sparse representation and dictionary learning, we will transform the deep network structure based on the aforementioned multi-attribute sparse representation model. The relevant model is shown in formula (14).

[0124]

[0125] On the one hand, we use deep dictionary learning to replace single shallow dictionary learning. Our goal is to learn the global structural characteristics that can represent the attributes of each image in the early dictionary learning layer. That is, the early stage focuses on the coverage expression of the overall structure, while the later dictionary learning layer focuses on the refined learning of local detail information, further optimizing the similarity relationship between sparse coefficients in the local space, making the dictionary more discriminative, and obtaining more accurate sparse coding. At the same time, the deep dictionary learning structure is embedded in the sparse representation model of multi-attribute component analysis, and an integrated learning network combining sparse representation and deep dictionary learning is constructed, so that the learned dictionary of each layer can be fed back to the sparse representation of various attribute features in a timely manner, forming an alternating optimization process of sparse representation and dictionary learning. The basic framework of the specific implementation is as follows: Figure 7 As shown;

[0126] The above description is only a further embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and concepts of the present invention within the scope disclosed by the present invention, which belong to the protection scope of the present invention.

Claims

1. A remote sensing image classification algorithm for multi-attribute feature extraction, characterized by: The steps include: Step 1: Extract sub-blocks from the image as the source of dictionary learning; Step 2: Combine deep dictionary learning on the selected sub-tile set to construct the corresponding dictionary basis for each attribute category; Step 3: Based on sparse coding technology, realize sparse decomposition of attribute components and solve sparse expression coefficients.

2. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 1, characterized in that: In step 2, the learning and construction of the dictionary includes content, roughness, contrast, and directionality.

3. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 2 is characterized in that: The roughness is used to characterize the edge characteristics in the image; Component separation based on roughness attribute, the image is separated into coarse component and fine component, where the coarse component is used to extract the large-scale strong texture edge features in the image, and the fine component is used to extract the small-scale weak texture edge features in the image; For the dictionary of coarse components, bilateral filtering technique is used to generate it; The specific formula is shown in (3): Among them, g(i,j) represents the output pixel at the spatial coordinate (i,j); f(k,l) represents the pixel at its neighborhood position (k,l); ω(i,j,k,l) ​​represents the weight coefficient; ∑ k,l Indicates the weighted sum of all neighboring pixels; When the bilateral filtering technique is applied to an image, the intensity value of each image pixel is calculated by the weighted average of the intensity values ​​of other pixels in its neighborhood. The weight is determined using Gaussian distribution, as shown in formula (4): Right now, The weight coefficient ω(i,j,k,l) ​​depends on the product of the domain kernel and the range kernel; σ d represents the variance of the domain kernel; σ r represents the variance of the range kernel; (i, j) represents the image pixel position coordinates; (k,l) represents the coordinates of the pixel position in its neighborhood; f(i,j) is the intensity value of the image pixel; f(k,l) represents the intensity value of the neighboring pixel; Wavelet threshold filtering technology is used to extract fine components. The specific formula (5) is as follows: Among them, H(x,y) is the coefficient of the horizontal subband at (x,y) after wavelet threshold filtering; h(x,y), v(x,y), and d(x,y) correspond to the coefficients of the horizontal, vertical, and diagonal subbands at (x,y) before wavelet transform, respectively; T h , T ν and T d are the thresholds in different directions respectively.

4. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 3 is characterized in that: Contrast uses the rate of change of pixel color brightness values ​​within a local neighborhood window to describe the contrast attribute; Component separation based on contrast attribute, the image is separated into two components: high contrast and low contrast, which represent the speed of the change rate of color brightness value respectively; Anisotropic diffusion transform is used to extract contrast component features, and its basic formula (6) is as follows: Among them, △ is the Laplace operator; is the gradient operator; div is the divergence; c(x,y,t) is the diffusion coefficient, which is used to control the diffusion rate; The two diffusion coefficient equations proposed by Perona and Malik are expressed as: Where K is a constant used to control the sensitivity of the edge; It means to find the gradient of I; represents the diffusion coefficient for calculating the gradient; K is the edge sensitivity control constant.

5. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 4 is characterized in that: Directionality is used to statistically describe the directional information of texture in an image; Select two orthogonal directional attributes, horizontal and vertical, and decompose the image into two attribute components in the horizontal and vertical directions by designing dictionary descriptions of two specific directions; The threshold filtering technology based on stationary wavelet transform is used to filter in a specific direction. The specific formula is as follows: Among them, ω j,k is the high frequency subband coefficient of each level of stationary wavelet transform; T is the set wavelet threshold; ω h Represents the coefficients of the stationary wavelet transform in the horizontal direction; T h Indicates the threshold in the horizontal direction; If ω h The absolute value of T h , the high-frequency subband coefficient is thresholded to 0, otherwise, a wavelet threshold constant is subtracted from the subband coefficient and multiplied by the corresponding sign function sign().

6. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 2 is characterized in that: It also includes linelikeness, regularity, and roughness, which are applied to the multi-attribute component analysis method by designing the corresponding dictionary generation algorithm.

7. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 6 is characterized in that: In step 3, the solution involves setting A s and A t The dictionary corresponding to the given texture attribute features, x s and x t They are the smoothing components y s and detail components t Sparse expression coefficients under the corresponding dictionary; For a given image y, the following optimization equation is given: y=y s +y t +n=A s x s +A t x t +n (10); It can consider different categories of texture attribute features and learn to construct corresponding dictionaries A according to different image attributes. s and A t , generate image components y that represent different texture attribute characteristics s and t ; That is, expand (10) to obtain the constrained optimization solution formula (11): Among them, λ1 and λ2 are regularization parameters; T s =(A T s A s ) -1 A T s and T t =(A T t A t ) -1 A T t Yes A s and A t The pseudo-inverse matrix of y s =A sxs and t =A txt Export; In the sparse coding stage, Equation (11) is solved using the variational and augmented Lagrangian algorithms. s and A t A deep sparse dictionary learning method with total variation and hard threshold regularization constraints is used to iteratively update the relevant dictionary.

8. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 7, characterized in that: Extraction and separation of multi-attribute component features, design of adaptive multi-attribute dictionary, so that each attribute feature has sparse expression under its own dictionary, but not sparse under other attribute dictionaries, so as to achieve the discrimination and separation of different attribute features under the multi-attribute joint dictionary representation. The relevant model formula is as follows: In formula (12), the original image S i Separated into linear combinations of different attribute component features, and α ik Then it represents the attribute component s ik In its corresponding dictionary D k The sparse representation coefficient under ; Therefore, the sparse representation model of attribute component analysis is used to learn multi-attribute features, which is expressed in the form of formula (13): Where, D: = (D1, D2, ..., D K ) represents a multi-attribute joint dictionary, Θ(D) is used to characterize the compatibility and discriminative regularity between attribute dictionaries; D is a joint dictionary, where D1, D2, etc. represent dictionaries of certain attributes; α represents the joint sparse coding coefficient of the original image under the joint dictionary D; Indicates the calculation of the p-norm of the matrix; Θ(D) is a discriminant function that characterizes the discriminability of each attribute dictionary in dictionary D; Formula (13) further includes the following steps to separate each attribute feature: The initial dictionary base is expanded by using spectral bundles or multi-endmember spectral sets, and the spectral heterogeneity is adaptively expressed by using multi-endmember spectral sets. The learning problem of multi-attribute features is converted into a class-level sparse representation reconstruction problem, and the class-level dictionary is optimized by using the image sub-block set corresponding to each attribute feature. After obtaining the dictionary of each attribute feature, a sparse representation model of multiple attribute features is constructed under the joint sparse representation of multiple attribute dictionaries. For multiple attribute features, a common dictionary between attribute features and a personalized dictionary for each attribute feature are constructed. The semantic connection between multiple attribute features establishes an abundance multi-attribute sparse representation model.

9. The remote sensing image classification algorithm for multi-attribute feature extraction according to claim 8, characterized in that: Combining the deep learning model with sparse representation and dictionary learning, the deep network structure is transformed based on the aforementioned multi-attribute sparse representation model. The relevant model is shown in formula (14): D1, D2, ...DN represent dictionaries of certain attributes; α represents the joint sparse coding coefficient of the original image under the joint dictionary D; Indicates the calculation of the p-norm of the matrix; F represents the Fibonacci norm of the matrix; Θ(D) is a discriminant function that represents the dictionary D i The discriminability of Φ() is a function used to regularize the result of the product of the dictionary and the coefficients.

Images