A regional subspace fusion method for infrared image and GM-APD image

Abstract

The application is a regional subspace fusion method for infrared images and GM-APD images. The application relates to the technical field of image signal processing. The application obtains a reference segmentation label through a laser image and an infrared image; performs pre-fusion on an unsegmented feature map and a segmented feature map; and finally fuses the segmented fusion map and the unsegmented fusion map. The application adds a regional factor in the whole fusion process, breaks through the limitation that the traditional double-mode whole image feature fusion cannot distinguish regions, and causes differences between regions and insufficient use of information within regions, so as to maximize the target saliency of infrared and the separability between GM-APD targets, successfully solves the problem that similar targets and target backgrounds are difficult to distinguish in the fusion process. This has important significance for realizing accurate target recognition of all-weather detection.

Description

Technical Field

[0001] This invention relates to the field of image signal processing technology, and is a regional subspace fusion method for infrared images and GM-APD images. Background Technology

[0002] In recent years, with the continuous development of technology, the requirements for multi-target recognition have become increasingly stringent. Single detectors offer limited information acquisition, making it difficult to meet practical needs. Multi-detector fusion technology has emerged and attracted widespread attention in both military and civilian fields. GM-APD possesses single-photon detection capability and extremely high range detection accuracy, effectively distinguishing multiple targets at different distances. However, it is susceptible to noise suppression and gating limitations, leading to incomplete target detection and missed detections. Infrared detection offers high resolution and strong anti-interference capabilities, enabling the detection of relatively complete targets in complex scenes. It reflects the target's thermal radiation information, exhibiting a certain degree of target-background separability. However, similar targets or targets and backgrounds at certain distances may exhibit similar thermal radiation characteristics, potentially leading to indistinguishability. Therefore, neither infrared images nor GM-APD images alone can achieve complete detection and effective differentiation of multiple targets. However, the complementary nature of the two modes allows for their fusion to effectively distinguish multiple targets.

[0003] Existing feature-level fusion algorithms are relatively few, mainly falling into three categories: one is used as part of deep learning, with concatenation or simple superposition as input; another is used directly as a module in deep learning training, but often requires a large number of training samples, posing a significant obstacle to practical applications; and the third focuses on mathematical reasoning, training PCA, CCA, or DCA prediction models to perform data multiplication or compression to increase data separability, but these algorithms are generally more effective for linearly separable cases and perform poorly in complex real-world scenarios. Furthermore, infrared images exhibit similarity among similar targets, and GM-APD images suffer from missed detections, making it difficult to effectively distinguish multiple targets using these methods directly. Meanwhile, comparing the advantages of dual-mode, it can be found that GM-APD images have differences between targets, while infrared images have differences between background and targets. These advantages can be superimposed through regional processing. At the same time, regional processing has a certain degree of noise isolation. If different targets and target-background can be segmented in the early stage of fusion, and then the similarity within the region and the differences between regions are added to the entire fusion process, the advantages of dual-mode can be added to the fusion, enhancing the separability between targets and reducing the misclassification within targets. Summary of the Invention

[0004] This invention addresses the problem that most existing multi-modal feature fusion methods treat the entire image as a whole, ignoring the characteristics and structural differences of different target regions, resulting in insufficient utilization of target region features and failure to fully leverage the advantages of fusion between GM-APD and infrared images. It proposes a regional subspace fusion algorithm for infrared and GM-APD images, aiming to maximize the target saliency in infrared images and the separability between GM-APD targets.

[0005] This invention provides the following technical solutions: Step 1: Obtain reference segmentation labels using laser and infrared images; Step 2: Pre-fuse the unsegmented feature maps and segmented feature maps; Step 3: Perform the final fusion of the segmented fused image and the unsegmented fused image.

[0006] Preferably, step 1 specifically comprises: Step 1.1: Extract feature maps of R and L by performing multi-scale hybrid feature map extraction on the laser image L and the infrared image R; Step 1.2: Perform superpixel segmentation on the infrared image R, and use the superpixel segmentation map to segment the feature maps of R and L, and combine them to obtain Superpixel segmented Maps; Step 1.3: Evaluate the region complexity of the Superpixel segmented Maps, perform background merging on the superpixel segmentation results, and obtain the segmentation reference label Re.

[0007] Preferably, the formula for calculating the region complexity is as follows: (1) Where p(i) is the probability of the i-th gray level appearing, i.e., the proportion of pixels of that gray level to the total number of pixels; p(i,j) is the probability that the i-th gray level and the j-th gray level appear simultaneously in the gray-level co-occurrence matrix; M is the total number of gray levels; |B b | represents the number of pixels in the b-th region; n(i) is the number of pixels in the i-th gray level; Perform the above operation on each superpixel segmentation block b of each feature map in the Superpixel segmented Maps, and then average them to obtain the final region complexity of the superpixel segmentation block b. The calculation formula is as follows: (2) Among them, C (b,k) T is the region complexity of the b-th superpixel segmentation block of the k-th feature map. bThis is the total number of feature maps in Superpixelsegmented Maps.

[0008] Preferably, when C (b) If the value is less than the threshold, the b region corresponding to the superpixel segmentation result is changed to the background. The above operation is performed on all superpixel regions to obtain the final background merging segmentation reference image Re.

[0009] Preferably, step 2 specifically comprises: Multi-scale hybrid feature map extraction is performed on the laser image L and the infrared image R to obtain feature maps of R and L. Then, subspace pre-fusion is performed on the unsegmented feature map and the segmented feature map of the reference Re.

[0010] Preferably, step 3 specifically comprises: Step 3.1: Perform a low-rank subspace representation on the two pre-fused images, as shown below. (3) Where S represents the segmented fused graph, U represents the unsegmented fused graph, q1 and q2 represent its noise and model error respectively, N1 and N2 represent its subspace projection matrix, and F represents the final fused feature. Step 3.2: The objective function for the fusion of region-constrained subspaces is proposed as follows: Minimize the following objective function to estimate the fused feature F: (4) Among them, the constraint terms are to ensure the regularity of the subspace; (5) in, α The weights are represented as the weights of the segmented fused graph and the unsegmented fused graph. β This represents the weights of the TV regular expression and the IN regular expression; The region regularization equation is shown below: (6) Among them, || || IN This represents the regional difference regularization; B represents the total number of segmented regions. b Let |B| represent the set of pixels in the b-th region; b | represents the number of pixels in the b-th region; F (i) and F (i') : represent the th element in the fused feature matrix F, respectively. i The and the first i The feature vector at ' pixels; Step 3.3: Use ADMM to solve the IN regularization part of the equation. Before each TV regularization, first perform || || on F. IN Processing Then, perform an overall iterative solution, transforming the solution for F into a solution for... The solution can be transformed into the following: (7) (8) in, Given the updated F, the subsequent solution is... Solve for; The TV regularization part of the equation is decomposed, and a penalty method is used to apply it. By constraining G, we can transform the problem into an augmented minimization problem: (9) Where K is the Lagrange operator and η is the penalty parameter.

[0011] Preferably, at the beginning of the iteration, the auxiliary variable G and the Lagrange multiplier K are both initialized to 0.

[0012] A regional subspace fusion system for infrared images and GM-APD images, the system comprising: The tag module obtains reference segmentation tags through laser images and infrared images; The pre-fusion module performs pre-fusion on the unsegmented feature map and the segmented feature map; The final fusion module performs a final fusion of the segmented fusion map and the unsegmented fusion map.

[0013] A computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a regional subspace fusion method for infrared images and GM-APD images.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement a regional subspace fusion method for infrared images and GM-APD images.

[0015] The present invention has the following beneficial effects: This invention incorporates a regional factor approach throughout the fusion process, overcoming the limitations of traditional dual-mode whole-image feature-indiscriminate fusion, which leads to regional differences and insufficient utilization of information within regions. This maximizes the saliency of infrared targets and the separability between GM-APD targets, successfully solving the problem of difficulty in distinguishing similar targets and target-background information during the fusion process. This is of great significance for achieving accurate target identification in all-weather, all-day detection. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 The diagram shown is a flowchart illustrating the construction of the reference segmentation label of the present invention; Figure 2 The flowchart shown is a pre-fusion process of the unsegmented feature map and the segmented feature map of the present invention; Figure 3 The image shows a comparison of the KNN classification results based on different feature maps according to the present invention. Figure 4 The image shows a comparison of the classification results of different methods of this invention. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The present invention will be described in detail below with reference to specific embodiments. Specific Implementation Example 1: according to Figures 1 to 4 As shown, the specific optimization technical solution adopted by the present invention to solve the above-mentioned technical problems is: The present invention relates to a regional subspace fusion method for infrared images and GM-APD images.

[0021] A method for region subspace fusion of infrared images and GM-APD images, the method comprising the following steps: Step 1: Obtain reference segmentation labels using laser and infrared images; Step 1 specifically involves: Step 1.1: Extract feature maps of R and L by performing multi-scale hybrid feature map extraction on the laser image L and the infrared image R; Step 1.2: Perform superpixel segmentation on the infrared image R, and use the superpixel segmentation map to segment the feature maps of R and L, and combine them to obtain Superpixel segmented Maps; Step 1.3: Evaluate the region complexity of the Superpixel segmented Maps, perform background merging on the superpixel segmentation results, and obtain the segmentation reference label Re.

[0022] The formula for calculating domain complexity is as follows: (1) Where p(i) is the probability of the i-th gray level appearing, i.e., the proportion of pixels of that gray level to the total number of pixels; p(i,j) is the probability that the i-th gray level and the j-th gray level appear simultaneously in the gray-level co-occurrence matrix; M is the total number of gray levels; |B b | represents the number of pixels in the b-th region; n(i) is the number of pixels in the i-th gray level; Perform the above operation on each superpixel segmentation block b of each feature map in the Superpixel segmented Maps, and then average them to obtain the final region complexity of the superpixel segmentation block b. The calculation formula is as follows: (2) Among them, C (b,k) T is the region complexity of the b-th superpixel segmentation block of the k-th feature map. b This is the total number of feature maps in Superpixelsegmented Maps.

[0023] When C (b) If the value is less than the threshold, the b region corresponding to the superpixel segmentation result is changed to the background. The above operation is performed on all superpixel regions to obtain the final background merging segmentation reference image Re.

[0024] Step 2: Pre-fuse the unsegmented feature maps and segmented feature maps; Step 2 specifically involves: Multi-scale hybrid feature map extraction is performed on the laser image L and the infrared image R to obtain feature maps of R and L. Then, subspace pre-fusion is performed on the unsegmented feature map and the segmented feature map of the reference Re.

[0025] Step 3: Perform the final fusion of the segmented fused image and the unsegmented fused image.

[0026] Step 3 specifically involves: Step 3.1: Perform a low-rank subspace representation on the two pre-fused images, as shown below. (3) Where S represents the segmented fused graph, U represents the unsegmented fused graph, q1 and q2 represent its noise and model error respectively, N1 and N2 represent its subspace projection matrix, and F represents the final fused feature. Step 3.2: The objective function for the fusion of region-constrained subspaces is proposed as follows: Minimize the following objective function to estimate the fused feature F: (4) Among them, the constraint terms are to ensure the regularity of the subspace; (5) in, α The weights are represented as the weights of the segmented fused graph and the unsegmented fused graph. β This represents the weights of the TV regular expression and the IN regular expression; The region regularization equation is shown below: (6) Among them, || || IN This represents the regional difference regularization; B represents the total number of segmented regions. b Let |B| represent the set of pixels in the b-th region; b | represents the number of pixels in the b-th region; F (i) and F (i') : represent the th element in the fused feature matrix F, respectively. i The and the first i The feature vector at ' pixels; Step 3.3: Use ADMM to solve the IN regularization part of the equation. Before each TV regularization, first perform || || on F. IN Processing Then, perform an overall iterative solution, transforming the solution for F into a solution for... The solution can be transformed into the following: (7) (8) in, Given the updated F, the subsequent solution is... Solve for; The TV regularization part of the equation is decomposed, and a penalty method is used to apply it. By constraining G, we can transform the problem into an augmented minimization problem: (9) Where K is the Lagrange operator and η is the penalty parameter.

[0027] At the beginning of the iteration, the auxiliary variable G and the Lagrange multiplier K are initialized to 0.

[0028] The present invention also provides a regional subspace fusion system for infrared images and GM-APD images, characterized in that: the system comprises: The tag module obtains reference segmentation tags through laser images and infrared images; The pre-fusion module performs pre-fusion on the unsegmented feature map and the segmented feature map; The final fusion module performs a final fusion of the segmented fusion map and the unsegmented fusion map.

[0029] The present invention provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a regional subspace fusion method for infrared images and GM-APD images.

[0030] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a regional subspace fusion method for infrared images and GM-APD images. Specific Implementation Example 2: The purpose of this invention is to address the problem that most existing multi-modal feature fusion methods treat the entire image as a whole, ignoring the characteristics and structural differences of different target regions, resulting in insufficient utilization of target region features and failure to fully leverage the advantages of fusion between GM-APD images and infrared images. This invention proposes a regional subspace fusion algorithm for infrared images and GM-APD images, aiming to maximize the target saliency in infrared images and the separability between GM-APD targets.

[0032] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps: Step 1: Construction of Reference Segmentation Labels (Re) Reference segmentation labels are obtained using laser and infrared images, as shown in the flowchart below. Figure 1 As shown: The flowchart for Step 1 is shown below. The main steps are: (1) Extract the feature maps of R and L by multi-scale hybrid feature maps (feature maps of DLatLRR, phase consistency feature map, and 2-D-SSA feature map of three scales) from the laser image (L) and infrared image (R).

[0033] (2) Perform superpixel segmentation on the original infrared image (R), and use the superpixel segmentation map to segment the feature maps of R and L, and combine them to obtain Superpixel segmented Maps.

[0034] (3) The region complexity of the Superpixel segmented Maps is evaluated, and then the background merging process is performed on the superpixel segmentation results to obtain the segmentation reference label (Re).

[0035] The formula for calculating domain complexity is as follows: (1) Where p(i) is the probability of the i-th gray level appearing, that is, the proportion of the number of pixels of this gray level to the total number of pixels; p(i,j) is the probability of the i-th gray level and the j-th gray level appearing at the same time in the gray-level co-occurrence matrix (GLCM); M is the total number of gray levels; |Bb| represents the number of pixels in the b-th region; n(i) is the number of pixels of the i-th gray level.

[0036] The first term of formula (1) is the formula for calculating the entropy of the gray histogram, and the second to fifth terms are the formulas for calculating the contrast, homogeneity, energy, and object complexity of the gray co-occurrence matrix (GLCM).

[0037] Perform the above operation on each superpixel segmentation block b of each feature map in the Superpixel segmented Maps, and then average them to obtain the final region complexity of the superpixel segmentation block b. The calculation formula is as follows: (2) Among them, C (b,k) T is the region complexity of the b-th superpixel segmentation block of the k-th feature map. b This is the total number of feature maps in Superpixelsegmented Maps.

[0038] If C (b) If the value is less than the threshold, the b region corresponding to the superpixel segmentation result is changed to the background. The above operation is performed on all superpixel regions to obtain the final background merging segmentation reference map (Re).

[0039] Step 2: Pre-fusion of unsegmented feature maps and segmented feature maps Multi-scale hybrid feature map extraction is performed on the laser image (L) and the infrared image (R) to obtain feature maps of R and L. Then, subspace pre-fusion is performed on the unsegmented feature map and the segmented feature map of the reference Re. The flowchart is as follows. Figure 2 As shown: Step 3: Dual Regularization Fusion The obtained segmented fused map (S) and unsegmented fused map (U) are then fused to obtain the final fused map (F) with dual-mode global features and region image features. The specific steps are as follows: 1. Subspace Representation Innovation: This invention proposes a Region Subspace Fusion (RSF) algorithm for infrared images and GM-APD images. By adding regional factors throughout the fusion process, it overcomes the limitations of traditional dual-mode whole-image feature-indiscriminate fusion, which leads to differences between regions and insufficient utilization of information within regions. This maximizes the saliency of infrared targets and the separability between GM-APD targets, successfully solving the problem of difficulty in distinguishing similar targets and target backgrounds during the fusion process.

[0040] First, the two pre-fused images are represented in a low-rank subspace as follows: (3) Where S represents the segmented fusion graph, U represents the unsegmented fusion graph, q1 and q2 represent its noise and model error respectively, N1 and N2 represent its subspace projection matrix, and F represents the final fusion feature.

[0041] 2. Feature Fusion The ultimate goal of fusion is to preserve both the overall and regional features of the two fused images while preventing overfitting. Therefore, a region-discrepancy regularization is proposed and combined with the TV regularization to form a joint constraint. This joint constraint is then integrated with the subspace fusion equation, resulting in the objective function for region-constrained subspace fusion, as shown below: Minimize the following objective function to estimate the fused feature F: (4) The constraint term is used to ensure the regularity of the subspace.

[0042] (5) in, α The weights are represented as the weights of the segmented fused graph and the unsegmented fused graph. β This represents the weights of the TV and IN regular expressions. Here, since the essence of the TV and IN regular expressions is to constrain the overall features based on smoothness and regional differences, they can be considered as a unified constraint. Therefore, a joint constraint strategy is adopted, and the parameters of the two regular expressions are... β and (1- β Constraints are used to ensure the orthogonality of two sub-functions.

[0043] The TV regular expression, || || TV The region regularization equation, representing the smoothing function between pixels, is shown below: (6) Among them, || || IN This represents the regional difference regularization; B represents the total number of segmented regions. b Let |B| represent the set of pixels in the b-th region; b | represents the number of pixels in the b-th region; F (i) and F (i') : represent the th element in the fused feature matrix F, respectively. i The and the first i The feature vector at ' pixels.

[0044] 3. Solving the equations The above formula (5) is solved using ADMM.

[0045] First, the IN regularization part of the equation is solved. To ensure the effectiveness of the dual regularization joint constraint and simplify the solution, the IN regularization part is combined with the TV regularization calculation; that is, before each TV regularization, F is first calculated using || || IN Processing Then, an overall iterative solution is performed, transforming the solution for F into a solution for... Solving this equation not only reduces the number of parameters required to solve the objective equation, but also ensures that the final fusion equation includes the IN regularization result. Furthermore, the iterative fusion result at each step is subject to the joint constraints of the TV regularization and the IN regularization. Equation (5) is transformed into the following: (7) (8) in, Given the updated F, the subsequent solution is... Solve for it.

[0046] Then, the TV regularization part of the equation is decomposed, and a penalty method is used to apply it. By constraining G, we can transform the problem into an augmented minimization problem: (9) Here, K is the Lagrange operator, and η is the penalty parameter. It can be solved using optimization methods. At the beginning of the iteration, the auxiliary variable G and the Lagrange multiplier K are initialized to 0.

[0047] Figure 3 A comparison of KNN classification results based on different feature maps, in the following order: infrared image, GM-APD image, classification label image, and classification result image based on the KNN algorithm. (The classification result obtained by this invention is based on the feature map.) By comparing the KNN classification results of the feature fusion algorithm of this invention with those of the more advanced feature fusion algorithms (fusion technology based on principal component analysis and multi-scale two-dimensional spectral decomposition of segmented principal component analysis domain (MSF) and hyperspectral band selection clustering method based on region-aware latent feature fusion (RLFFC)), the classification improvement effect of the algorithm of this invention is analyzed, and Table 1 shows its objective evaluation index.

[0048] Table 1. Comparison of classification results for different methods

[0049] Figure 4 The images, in order, are infrared images, GM-APD images, classification-labeled images, and classification results based on the MSF algorithm, RLFFC algorithm, and the algorithm of this invention.

[0050] From Table 1 and Figure 4 As shown, when classification is performed based on the results of the MSF algorithm, there are a large number of misclassifications between targets. When classification is performed based on the results of the RLFFC algorithm, there are also a large number of misclassifications between targets and background misclassifications as targets. However, after adding the region factor, the algorithm in this paper basically solves the problem of background misclassification as targets, and the targets are detected relatively completely, with only sporadic misclassifications between targets. The classification improvement effect is very significant.

[0051] The above description is merely a preferred embodiment of a regional subspace fusion method for infrared images and GM-APD images. The scope of protection for such a method is not limited to the above embodiments; all technical solutions falling within this conceptual framework are within the protection scope of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the protection scope of this invention.

Claims

1. A method for region subspace fusion of infrared images and GM-APD images, characterized by: The method includes the following steps: Step 1: Obtain reference segmentation labels using laser and infrared images; Step 2: Pre-fuse the unsegmented feature maps and segmented feature maps; Step 3: Perform the final fusion of the segmented fused image and the unsegmented fused image.

2. The method according to claim 1, characterized in that: Step 1 specifically involves: Step 1.1: Extract feature maps of R and L by performing multi-scale hybrid feature map extraction on the laser image L and the infrared image R; Step 1.2: Perform superpixel segmentation on the infrared image R, and use the superpixel segmentation map to segment the feature maps of R and L, and combine them to obtain Superpixel segmented Maps; Step 1.3: Evaluate the region complexity of the Superpixel segmented Maps, perform background merging on the superpixel segmentation results, and obtain the segmentation reference label Re.

3. The method according to claim 2, characterized in that: The formula for calculating domain complexity is as follows: （1） Where p(i) is the probability of the i-th gray level appearing, i.e., the proportion of pixels of that gray level to the total number of pixels; p(i,j) is the probability that the i-th gray level and the j-th gray level appear simultaneously in the gray-level co-occurrence matrix; M is the total number of gray levels; |B b | represents the number of pixels in the b-th region; n(i) is the number of pixels in the i-th gray level; Perform the above operation on each superpixel segmentation block b of each feature map in the Superpixel segmented Maps, and then average them to obtain the final region complexity of the superpixel segmentation block b. The calculation formula is as follows: （2） Among them, C (b,k) T is the region complexity of the b-th superpixel segmentation block of the k-th feature map. b This is the total number of feature maps in Superpixelsegmented Maps.

4. The method according to claim 3, characterized in that: When C (b) If the value is less than the threshold, the b region corresponding to the superpixel segmentation result is changed to the background. The above operation is performed on all superpixel regions to obtain the final background merging segmentation reference image Re.

5. The method according to claim 4, characterized in that: Step 2 specifically involves: Multi-scale hybrid feature map extraction is performed on laser image L and infrared image R to obtain feature maps of R and L. Subspace pre-fusion is then performed on the unsegmented feature map and the segmented feature map of reference Re.

6. The method according to claim 5, characterized in that: Step 3 specifically involves: Step 3.1: Perform a low-rank subspace representation on the two pre-fused images, as shown below. （3） Where S represents the segmented fused graph, U represents the unsegmented fused graph, q1 and q2 represent its noise and model error respectively, N1 and N2 represent its subspace projection matrix, and F represents the final fused feature. Step 3.2: The objective function for the fusion of region-constrained subspaces is proposed as follows: Minimize the following objective function to estimate the fused feature F: （4） Among them, the constraint terms are to ensure the regularity of the subspace; （5） in, α The weights are represented as the weights of the segmented fused graph and the unsegmented fused graph. β This represents the weights of the TV regular expression and the IN regular expression; The region regularization equation is shown below: （6） Among them, || || IN This represents the regional difference regularization; B represents the total number of segmented regions. b Let |B| represent the set of pixels in the b-th region; b | represents the number of pixels in the b-th region; F (i) and F (i') : represent the th element in the fused feature matrix F, respectively. i The and the first i The feature vector at ' pixels; Step 3.3: Use ADMM to solve the IN regularization part of the equation. Before each TV regularization, first perform || || on F. IN Processing obtained Then, perform an overall iterative solution, transforming the solution for F into a solution for... The solution can be transformed into the following: （7） (8) in, Given the updated F, the subsequent solution is... Solving for; The TV regularization part of the equation is decomposed, and a penalty method is used to apply it. By constraining G, we can transform the problem into an augmented minimization problem: (9) Where K is the Lagrange operator and η is the penalty parameter.

7. The method according to claim 6, characterized in that: At the beginning of the iteration, the auxiliary variable G and the Lagrange multiplier K are initialized to 0.

8. A regional subspace fusion system for infrared images and GM-APD images, characterized in that: The system includes: The tag module obtains reference segmentation tags through laser images and infrared images; The pre-fusion module performs pre-fusion on the unsegmented feature map and the segmented feature map; The final fusion module performs a final fusion of the segmented fusion map and the unsegmented fusion map.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method as claimed in any one of claims 1-7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the method of any one of claims 1-7.

Images