A night large field of view video image small target change detection method and device

Abstract

The application discloses a night large field of view video image small target change detection method and device, the method comprises the following steps: using comprehensive Anscombe root transformation and total variation regularization to remove the residual Poisson noise in the difference map, and obtaining the denoising image through inverse transformation; using total variation regularization based on partial differential equation and minimizing energy functional to remove the residual Gaussian noise in the denoising image; using the original dual algorithm to solve the energy functional; processing the difference map based on the multi-scale superpixel segmentation and reconstruction method of statistical information, which is used for enhancing the edge features of the change region and suppressing the interference of noise; fusing the two difference maps based on the multi-scale singular value decomposition fusion algorithm, and segmenting the fused difference map by using the FCM clustering algorithm to obtain the detection result. The device comprises a processor and a memory. The application fuses two difference maps by using the multi-scale singular value decomposition fusion algorithm to obtain a robust difference map, and finally obtains a change detection result with complete detail structure by using a clustering algorithm.

Description

Technical Field

[0001] This invention relates to the field of nighttime wide field-of-view video image changes, and more particularly to a method and apparatus for detecting small target changes in nighttime wide field-of-view video images. Background Technology

[0002] With the development of technology and the needs of social governance, surveillance equipment has become widely used. In conventional scenarios, video image change detection algorithms for detecting changes in surveillance video images are already very mature. However, in nighttime environments with large fields of view, surveillance video images suffer from complex random noise and low signal-to-noise ratios, making it difficult for people to detect small, changing targets. Therefore, researching change detection algorithms suitable for nighttime environments with large fields of view has become a very important topic.

[0003] Currently, research on nighttime wide-field-of-view video image processing mainly focuses on the acquisition of wide-field-of-view images and nighttime image enhancement, while research on change detection in nighttime wide-field-of-view video images is relatively insufficient. For change detection tasks, the biggest challenge is suppressing noise in the difference map. The difference map is generated by comparing multi-temporal images and represents the degree of change for each pair of corresponding pixels. Because the noise types in the difference maps of nighttime wide-field-of-view surveillance videos are diverse and complex, difference maps generated using traditional subtraction and ratio operators often fail to accurately reflect the trend of change. This poses a challenge to extracting effective change targets from noise-inundated difference maps.

[0004] To address the aforementioned issues, it is necessary to improve the quality of the difference map to better distinguish between changed and unchanged areas. Summary of the Invention

[0005] This invention provides a method and apparatus for detecting small target changes in video images under a large field of view at night. The invention enhances two denoised difference maps through adaptive superpixel reconstruction, effectively enhancing the edge features of the difference maps. Furthermore, it fuses the two difference maps using a multi-scale singular value decomposition fusion algorithm to obtain a robust difference map. Finally, a clustering algorithm is used to obtain change detection results with complete detailed structure. See the description below for details:

[0006] A first aspect is a method for detecting small target changes in video images under a large field of view at night, the method comprising:

[0007] The residual Poisson noise in the difference map is removed by combining Anscombe root transform and total variation regularization, and the denoised image is obtained by inverse transform.

[0008] The residual Gaussian noise in the denoised image is removed by using total variational regularization based on partial differential equations and minimizing the energy generalization function; the primal dual algorithm is used to solve the above energy generalization function.

[0009] A multi-scale superpixel segmentation and reconstruction method based on statistical information is used to process the difference map to enhance the edge features of the change region while suppressing the interference of noise.

[0010] The two difference maps are fused based on multi-scale singular value decomposition fusion, and the detection result is obtained by segmenting the fused difference map using the FCM clustering algorithm.

[0011] Among them, Anscomb e The forward and inverse transformations of the root transform are:

[0012]

[0013]

[0014] Among them, DI k For the original image, X k X is the image after the Anscombe transform. k - 1 This is the image after the Anscombe inverse transform.

[0015] The method of using the primal dual algorithm to solve the above energy generalization function is as follows:

[0016] The original variable and the dual variable p are updated alternately, and the updates are performed according to the gradient descent principle of the regularized model. The formula for updating the original variable is:

[0017]

[0018] Among them, I k+1 This is the updated image. Let I0 be the gradient of the image, λ be the Lagrange multiplier, and I0 be the original noisy image. The update formula for the Lagrange multiplier λ is:

[0019]

[0020]

[0021] Among them, I gradient p represents the magnitude of the average gradient of the image, reflecting the overall gradient level of the image. i，x and p i，y These are the horizontal and vertical components of the gradient vector p, corresponding to the gradients in the x and y directions of the image, respectively.

[0022] Then, the dual variable p is updated. The update formula for the dual variable p is:

[0023]

[0024] Where σ is the step size parameter, It is the updated image gradient.

[0025] The multi-scale superpixel segmentation and reconstruction method based on statistical information processes the difference map as follows:

[0026] Given a set of pixel values ​​{x1, x2, ..., xn} within a superpixel block. n}, where n is the number of pixel values, and the reconstruction formula is:

[0027] I R =(w std ×O)+(w mean ×M)+(w median ×N)

[0028] Among them, w std w represents the weight of the original pixel. mean The weights representing the mean, w median The weights represent the median, where O is the original pixel value within the superpixel block, M is the mean within the superpixel block, and N is the median within the superpixel block.

[0029] The fusion of the two difference maps based on multi-scale singular value decomposition is as follows:

[0030] The selection rules for U and V for each layer are as follows:

[0031]

[0032]

[0033] For diagonal matrices, larger singular values ​​are selected for fusion, and the selection rule is as follows:

[0034] ∑ fused =max(∑1, ∑2)

[0035] For Contrast1 and Contrast2, μ1 is obtained by calculating the global standard deviation of different layers of the Laplacian pyramid. k and μ2 k It is the average value of the current layer image, where m and n are the number of rows and columns of the image, respectively;

[0036]

[0037]

[0038] The final fused image expression obtained through MSVD is:

[0039] Yk =U fused k ×∑ fused k ×(V fused k ) T

[0040]

[0041] Where K is the number of decomposition layers of the Laplace pyramid, and Y is the final fusion difference diagram.

[0042] In a second aspect, a device for detecting small target changes in video images under a wide field of view at night, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.

[0043] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.

[0044] The beneficial effects of the technical solution provided by this invention are:

[0045] 1. This invention employs a regularized denoising method with dynamically adjusted parameters to remove noise from the difference image. This method achieves a good balance between image detail preservation and denoising performance, and its denoising effect is superior to other methods.

[0046] 2. This invention designs a new superpixel reconstruction method to enhance the difference map, making the distinction between changing and invariant regions more obvious. The difference map obtained by this method can suppress noise and background and highlight changing pixels, reducing the probability of false detection.

[0047] 3. This invention proposes a novel difference graph fusion method based on multi-scale singular value decomposition. By selecting different fusion rules, this method can effectively overcome the influence of most random noise and capture the changing parts more prominently. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of a method for detecting small target changes in video images under a large field of view at night.

[0049] Figure 2 The flowchart is shown for the proposed change detection scheme.

[0050] Figure 3 A schematic diagram of the proposed MSVD fusion algorithm;

[0051] Figure 4 A schematic diagram of change detection results in a high-resolution video image;

[0052] Among them, (a) is the image of time phase 1; (b) is the image of time phase 2; and (c) is the change detection result after processing by this algorithm.

[0053] Figure 5 This is a schematic diagram of the change detection results for a low-resolution video image.

[0054] Among them, (a) is the image of time phase 1; (b) is the image of time phase 2; and (c) is the change detection result after processing by this algorithm. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0056] I. Denoising of Difference Graphs Based on Regularization

[0057] The quality of the difference map directly affects the final change detection result. However, due to noise, difference maps without further processing often fail to accurately reflect the true trend of change. For nighttime video images from eagle-eye surveillance cameras, the noise in the difference map mainly originates from the surveillance camera itself, including: dark current, shot noise following a Poisson distribution generated by illumination, and Gaussian noise with zero mean generated by sensor heat generation during imaging. In low-light conditions, due to the smaller number of photons, photon randomness becomes more significant, resulting in a greater impact of Poisson noise on the image.

[0058] Based on the above analysis, the embodiments of the present invention employ a comprehensive Anscomb... e Root transformation and total variational regularization methods are used to remove residual Poisson noise from the difference graph. (Anscomb) e Anscomb transform is a commonly used data transformation method in statistics. This method converts Poisson noise into approximately Gaussian noise with a stable variance, then uses a denoising method to remove additive Gaussian noise, and finally obtains the denoised image through inverse transform. e The formulas for the forward and inverse transformations are shown in equations (1) and (2):

[0059]

[0060]

[0061] Among them, DI k For the original image, X k X is the image after the Anscombe transform. k -1 This is the image after the Anscombe inverse transform.

[0062] Subsequently, this embodiment of the invention uses a total variation regularization denoising method based on partial differential equations to remove residual Gaussian noise in the difference map. The total variation model is an anisotropic model that relies on gradient descent flow to smooth the image, which can better preserve image edge texture and other details. The denoising principle is to minimize the energy generalization function, as shown in formula (3):

[0063]

[0064] The first term in the above formula The first term is the regularization term, which describes the smoothness of the image. The second term (λ∫|I-I0|) 2 dΩ) is the data fidelity term, I is the denoised image to be determined, and I0 is the original noisy image. The data fidelity term represents the similarity between the denoised image and the original image. The Lagrange multiplier λ is used to balance the regularization term and the data fidelity term.

[0065] Next, this embodiment of the invention uses the primal-dual algorithm to solve the above-mentioned energy generalization function. In each iteration, this embodiment of the invention alternately updates the primal variable and the dual variable p, and updates them according to the gradient descent principle of the regularized model. The specific formula for updating the primal variable is shown in (4):

[0066]

[0067] Among them, I k+1 This is the updated image. Let I0 be the gradient of the image, λ be the Lagrange multiplier, and I0 be the original noisy image. This embodiment of the invention incorporates a process of dynamically updating the Lagrange multiplier λ. λ is the average gradient value calculated based on the image gradient in the current iteration step. The value of λ is then adjusted by dividing the initial λ parameter by this average gradient value. The update formula for the Lagrange multiplier λ is:

[0068]

[0069]

[0070] Among them, I gradient p represents the magnitude of the average gradient of the image, reflecting the overall gradient level of the image. i，x and p i，y These are the horizontal and vertical components of the gradient vector p, corresponding to the gradients in the x and y directions of the image, respectively.

[0071] The dual variable p is then updated, and the update formula for the dual variable p is shown in equation (7):

[0072]

[0073] Where σ is the step size parameter, It is the updated image gradient.

[0074] II. Difference Map Enhancement Based on Adaptive Superpixel Reconstruction

[0075] After denoising the difference map, this embodiment of the invention uses a multi-scale superpixel segmentation and reconstruction method based on statistical information to further process the difference map, aiming to enhance the edge features of the change regions while further suppressing noise interference. First, the denoised difference map is segmented into multiple homogeneous superpixels using the SLIC (Simple Linear Iterative Clustering) superpixel segmentation method. Within each superpixel block, a weighted combination method is designed, using the original pixel values, mean, and standard deviation to weight the superpixel block for reconstruction. Different weights are assigned to each piece of information, and the weights are determined based on the standard deviation of the pixel values ​​within the superpixel block. A larger standard deviation indicates a more dispersed distribution of pixel values, meaning greater variability in pixel values; therefore, this embodiment of the invention aims to retain more of the original pixel values. Conversely, a smaller standard deviation indicates less variability in pixel values, and this embodiment of the invention tends to use the mean or median instead of the original pixel values. Based on the calculated weights, the pixel values ​​of the superpixel blocks are recalculated using a weighted average. This method can reduce the possible noise effects while preserving the original image information by using the mean and median.

[0076] Given a set of pixel values ​​{x1, x2, ..., xn} within a superpixel block. n}, where n is the number of pixel values. The reconstruction formula is shown in equation (8):

[0077] I R =(w std ×O)+(w mean ×M)+(w median ×N)#(8)

[0078] Among them, W std w represents the weight of the original pixel. mean w represents the weight of the mean. median The weights represent the median, where O is the original pixel value within the superpixel block, M is the mean within the superpixel block, and N is the median within the superpixel block.

[0079] Among them, wstd Used to reflect the degree of variation in pixel values ​​within a superpixel block; when the standard deviation is large, w std A value close to 1 indicates that the pixel values ​​within the superpixel block are more dispersed, and more of the original pixel values ​​will be retained. However, when the standard deviation σ is small, w... std The value will be close to 0, indicating that the pixel values ​​within the superpixel block are relatively concentrated. In this case, the mean and median will be considered more to reduce the possible noise effect.

[0080] In this embodiment of the invention, considering the high resolution of the image, four different segmentation scales were selected: 4000, 8000, 16000, and 32000. Image reconstruction was performed on the superpixel segmentation results at each scale. Finally, the reconstruction results at multiple scales were summed and averaged to obtain the final reconstructed image. This processing method comprehensively considers image details and features at different scales, thereby improving the quality and stability of image reconstruction. The formula for the final reconstruction result is:

[0081]

[0082] Where S is the number of segmentation scales, which is set to 4 in this embodiment of the invention. R t is the reconstructed image obtained by formula (9) at different scales.

[0083] III. Difference Graph Fusion Based on Multi-Scale Singular Value Decomposition Fusion

[0084] To address the issue of insufficient robustness of single-image difference maps, this embodiment of the invention uses a multi-scale singular value decomposition (MSVD) fusion method to fuse two difference maps. The MSVD algorithm is similar to wavelet transform, and its basic idea is to replace filtering with singular value decomposition at each level of the smoothing component. First, the original image is decomposed into images at multiple scales. In this embodiment of the invention, the difference map is decomposed into images at three different scales using the Laplacian pyramid. Then, singular value decomposition (SVD) is performed on the images at each scale. SVD is a matrix decomposition method that represents a matrix as the product of three matrices. The SVD decomposition process for the two Laplacian pyramid images at different scales is shown in (10) and (11):

[0085] L1 (k) =U1 (k) ∑1 (k) (V1 k ) T #(10)

[0086] L2 (k) =U2 (k) ∑2 (k)(V2 k ) T #(11)

[0087] Among them, L1 (k) and L2 (k) These are images of DI1 and DI2 at the k-th scale obtained through Laplacian pyramid sampling, where K is the total sampling scale, and U1 (k) and U2 (k) They are two orthogonal matrices, each containing L1 (k) and L2 (k) The left singular vector, ∑1 (k) and ∑2 (k) It is a diagonal matrix, (V1) k ) T and (V2) k ) T It is the transpose of two orthogonal matrices, each containing the image L1. (k) and L2 (k) The right singular vectors can be decomposed using SVD to obtain their respective basic spatial features. The diagonal matrix stores the main information of the image, while the left and right singular matrices store the structural information.

[0088] The selection rules for U and V for each layer are shown in equations (12) and (13). Since larger singular values ​​can better preserve the useful information of the original image, the embodiment of the present invention selects the larger singular values ​​for fusion for the diagonal matrix, and the selection rule is shown in equation (14).

[0089]

[0090]

[0091] ∑ fused =max(∑1,∑2)#(14)

[0092] For Contrast1 and Contrast2, the embodiments of the present invention obtain them by calculating the global standard deviation of different layers of the Laplace pyramid, μ1. k and μ2 k It is the average value of the current layer image, where m and n are the number of rows and columns of the image, respectively.

[0093]

[0094]

[0095] The final fused image expression obtained by MSVD is shown in equations (17) and (18):

[0096] Y k=U fused k ×∑ fused k ×(V fused k ) T #(17)

[0097]

[0098] Where K is the number of decomposition layers of the Laplace pyramid, and in this embodiment of the invention, K is 3, and Y is the final fusion difference map.

[0099] Finally, in this embodiment of the invention, the FCM clustering algorithm is used to segment the fused difference map to obtain the detection result.

[0100] experiment

[0101] To verify the effectiveness of this method in detecting changes in nighttime wide-field-of-view video images, experiments were conducted on nighttime wide-field-of-view video image datasets at two different resolutions. The datasets were captured multiple times at night using a camera fixed on a tripod in 25 different scenes. Two sets of images were selected from each scene, each consisting of two frames, resulting in 50 different image pairs. Each image pair initially had a resolution of 1920×1080. Subsequently, image downsampling was used to reduce the height and width of these 50 images to half their original size, resulting in another 50 image pairs with a resolution of 960×540. A total of 100 image pairs were obtained, with 50 sets at 1920×1080 resolution and the other 50 sets at 960×540 resolution.

[0102] Figure 1 To address the difficulty in detecting small, changing targets under wide field of view at night, this paper first employs an adaptive superpixel reconstruction method to reconstruct two denoised difference maps by selecting different segmentation parameters. Then, a multi-scale singular value fusion (SVF) method is used to fuse the two difference maps. SVF utilizes complementary information from different difference maps by selecting fusion rules at different scales to obtain a robust difference map. Finally, the final image of the changed object is obtained using the FCM clustering algorithm.

[0103] from Figure 4 and Figure 5It can be seen that this method can not only effectively detect changing targets, but also avoid blurring moving objects. False negatives (FN) (undetected changed pixels) and false positives (FP) (unchanged pixels mistakenly detected as changed) are calculated. Overall error (OE), classification accuracy (PCC), Kappa coefficient, and F1 score (F1) are also evaluated. PCC represents the classification precision, the Kappa coefficient assesses the consistency of the algorithm, and the F1 score is an important indicator for evaluating classification performance. This method uses these six objective indicators to measure the quality of the algorithm's detection results. From Table 1, which shows the evaluation indicators obtained by averaging 100 sets of experiments at each of the two resolutions, it is clear that this method performs better in terms of FN, FP, OE, PCC, Kappa, and F1 scores.

[0104] Table 1 Comparison of Objective Indicators

[0105]

[0106] A device for detecting small target changes in video images under a wide field of view at night, the device comprising: a processor and a memory, wherein the memory stores program instructions, and the processor calls the program instructions stored in the memory to cause the device to perform the following method steps:

[0107] Using Anscomb e Root transformation and total variation regularization remove residual Poisson noise in the difference map, and the denoised image is obtained by inverse transformation;

[0108] The residual Gaussian noise in the denoised image is removed by using total variational regularization based on partial differential equations and minimizing the energy generalization function; the primal dual algorithm is used to solve the above energy generalization function.

[0109] A multi-scale superpixel segmentation and reconstruction method based on statistical information is used to process the difference map to enhance the edge features of the change region while suppressing the interference of noise.

[0110] The two difference maps are fused based on multi-scale singular value decomposition fusion, and the detection result is obtained by segmenting the fused difference map using the FCM clustering algorithm.

[0111] Among them, Anscomb e The forward and inverse transformations of the root transform are:

[0112]

[0113]

[0114] Among them, DI k For the original image, X kX is the image after the Anscombe transform. k - 1 This is the image after the Anscombe inverse transform.

[0115] The energy generalization function described above is solved using the primal-dual algorithm as follows:

[0116] The original variable and the dual variable p are updated alternately, and the updates are performed according to the gradient descent principle of the regularized model. The formula for updating the original variable is:

[0117]

[0118] Among them, I k+1 This is the updated image. Let I0 be the gradient of the image, λ be the Lagrange multiplier, and I0 be the original noisy image. The update formula for the Lagrange multiplier λ is:

[0119]

[0120]

[0121] Among them, I gradient p represents the magnitude of the average gradient of the image, reflecting the overall gradient level of the image. i，x and p i，y These are the horizontal and vertical components of the gradient vector p, corresponding to the gradients in the x and y directions of the image, respectively.

[0122] Then, the dual variable p is updated. The update formula for the dual variable p is:

[0123]

[0124] Where σ is the step size parameter, It is the updated image gradient.

[0125] Among them, the multi-scale superpixel segmentation and reconstruction method based on statistical information processes the difference map as follows:

[0126] Given a set of pixel values ​​{x1, x2, ..., xn} within a superpixel block. n}, where n is the number of pixel values, and the reconstruction formula is:

[0127] I R =(w std ×O)+(w mean ×M)+(w median ×N)

[0128] Among them, W std w represents the weight of the original pixel.mean The weights representing the mean, w median The weights represent the median, where O is the original pixel value within the superpixel block, M is the mean within the superpixel block, and N is the median within the superpixel block.

[0129] Among them, the fusion of the two difference maps based on multi-scale singular value decomposition is as follows:

[0130] The selection rules for U and V for each layer are as follows:

[0131]

[0132]

[0133] For diagonal matrices, larger singular values ​​are selected for fusion, and the selection rule is as follows:

[0134] ∑ fused =ma(∑1, ∑2)

[0135] For Contrast1 and Contrast2, μ1 is obtained by calculating the global standard deviation of different layers of the Laplacian pyramid. k and μ2 k It is the average value of the current layer image, where m and n are the number of rows and columns of the image, respectively;

[0136]

[0137]

[0138] The final fused image expression obtained through MSVD is:

[0139] Y k =U fused k ×∑ fused k ×(V fused k ) T

[0140]

[0141] Where K is the number of decomposition layers of the Laplace pyramid, and Y is the final fusion difference diagram.

[0142] It should be noted that the device descriptions in the above embodiments correspond to the method descriptions in the embodiments, and the embodiments of the present invention will not be repeated here.

[0143] The execution entities of the aforementioned processor and memory can be devices with computing functions such as computers, microcontrollers, and single-chip microcomputers. In specific implementations, the embodiments of the present invention do not limit the execution entities and can select them according to the needs of actual applications.

[0144] Data signals are transmitted between the memory and the processor via a bus, which will not be elaborated upon in this embodiment of the invention.

[0145] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium, the storage medium including a stored program, which, when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiments.

[0146] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0147] It should be noted that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be repeated here.

[0148] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated.

[0149] A computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted through a computer-readable storage medium. A computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic or semiconductor, etc.

[0150] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.

[0151] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting small target changes in video images under a large field of view at night, characterized in that, The method includes: The residual Poisson noise in the difference map is removed by combining Anscombe root transform and total variation regularization, and the denoised image is obtained by inverse transform. Residual Gaussian noise in the denoised image is removed by using total variational regularization based on partial differential equations and minimizing the energy generalization function; the primal dual algorithm is used to solve the above energy generalization function. A multi-scale superpixel segmentation and reconstruction method based on statistical information is used to process the difference map to enhance the edge features of the change region while suppressing the interference of noise. The two difference maps are fused based on multi-scale singular value decomposition fusion, and the detection result is obtained by segmenting the fused difference map using the FCM clustering algorithm. The forward and inverse transformations of the Anscombe root transform are as follows: in, The original image. This is the image after the Anscombe transform. The image is after the inverse Anscombe transform; The statistical information-based multi-scale superpixel segmentation and reconstruction method processes the difference map as follows: Given a set of pixel values ​​within a superpixel block Where n is the number of pixel values, the reconstruction formula is: in, The weights of the original pixels, The weights representing the mean. The weights representing the median, It represents the original pixel values ​​within the superpixel block. It is the mean value within the superpixel block. It is the median value within the superpixel block; The fusion of the two difference maps based on multi-scale singular value decomposition is as follows: For each layer and The selection rules are as follows: in, and They are two orthogonal matrices. and They are two orthogonal matrices; For diagonal matrices, larger singular values ​​are selected for fusion, and the selection rule is as follows: For Contrast1 and Contrast2, the global standard deviation of different layers of the Laplace pyramid is calculated. and It is the average value of the current layer image, where m and n are the number of rows and columns of the image, respectively; The final fused image expression obtained through MSVD is: in, This represents the number of layers in the Laplace pyramid. This is the final fusion difference diagram.

2. The method for detecting small target changes in video images under a large field of view at night according to claim 1, characterized in that, The method of using the primal dual algorithm to solve the above energy generalization function is as follows: Alternately update the original variables and dual variables And update it according to the gradient descent principle of the regularized model, the original variable The updated formula is: in, This is the updated image. It is an image gradient, It is a Lagrange multiplier. It is the original noisy image, Lagrange multipliers The update formula is: in, It represents the magnitude of the average gradient of the image, reflecting the overall gradient level of the image. and It is the gradient vector The horizontal and vertical components correspond to the gradients in the x and y directions of the image, respectively. Then the dual variables Update the dual variable. The update formula is: ； in, It is the step size parameter. It is the updated image gradient.

3. A device for detecting small target changes in video images under a large field of view at night, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-2.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-2.

Images