Low-illumination wide-view-field video image change detection method and device

Abstract

The invention discloses a low-illumination wide-view-field video image change detection method and device, and the method comprises the steps: carrying out the SSR enhancement of a low-illumination image, and carrying out the principal component analysis and fusion of the enhanced image and a bright channel image; based on the enhanced low-illumination image, an improved arc tangent operator is adopted to generate a difference graph; in combination with chi-square transformation and an arc tangent operator, an improved multiplication fusion technology is proposed to process the difference chart, and a threshold value in the multiplication fusion technology is processed based on an approximation method of logarithmic normal distribution to obtain an optimized threshold value; and pixels in the difference image are processed based on fitting error minimization of logarithmic normal distribution, an optimal threshold segmentation difference image is obtained, and a change detection result is obtained. The device comprises a processor and a memory. The method is suitable for images with few change pixels and no change scene in a wide field of view, and the detection precision and robustness are improved.

Description

Technical Field

[0001] The present invention relates to the field of image change detection, and particularly to a method and device for detecting changes in low-light wide-field video images. Background Art

[0002] The main purpose of video image change detection is to detect changes between different video image frames collected by an eagle-eye monitoring device in the same scene. The eagle-eye monitoring device faces problems of a wide field of view and insufficient illumination under low-light conditions, resulting in a decline in image quality and an increased risk of false detection. [1] Currently, there is relatively little research on change detection of low-light enhanced wide-field video images. Existing research mainly focuses on enhancement and denoising of low-light images, [2] for example: based on joint filtering [3] , Retinex model [4] and so on. However, these methods have limitations in noise processing and illumination equalization, and are difficult to fully meet the requirements of low-light monitoring.

[0003] Traditional change detection methods, such as pixel-based, region analysis, and model classification, are widely used in remote sensing images, but they perform poorly in the face of complex noise problems under low-light conditions. In recent years, deep learning methods have gradually been applied to change detection tasks, such as using convolutional neural networks (CNNs) [5] and Transformer-based models [6] and so on. These methods have improved the detection accuracy by learning the deep features of images, but their requirements for large-scale labeled data and computing resources limit their practical applications.

[0004] References

[0005] [1] Wang J, Wang X, Li Y. Stray Light Nonuniform Background Elimination Method Based on Image Block Self-Adaptive Gray-Scale Morphology for Wide-Field Surveillance[J]. Applied Sciences, 2022, 12(14): 7299.

[0006] [2] Ma Q, Wang Y, Zeng T. Retinex-based variational framework for low-light image enhancement and denoising[J]. IEEE Transactions on Multimedia, 2022, 25: 5580-5588.

[0007] [3] Dong J, Pan J, Ren J S, et al. Learning spatially variant linear representation models for joint filtering[J]. IEEE transactions on pattern analysis and machine intelligence, 2021, 44(11): 8355-8370.

[0008] [4] Xu J, Hou Y, Ren D, et al. Star: A structure and texture aware retinex model[J]. IEEE Transactions on Image Processing, 2020, 29: 5022-5037.

[0009] [5] Jiang M, Chen Y, Dong Z, et al. Multiscale fusion CNN-transformer network for high-resolution remote sensing image change detection[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2024, 17: 5280-5293.

[0010] [6]Peng T,Hu L,Huang J,et al.A HRNet-Transformer Network CombiningRecurrent-Tokens for Remote Sensing Image Change Detection[C] / / ComputerGraphics International Conference.Cham:Springer Nature Switzerland,2023:15-26. Summary of the Invention

[0011] The present invention provides a method and device for detecting changes in low-light wide-field video images. The present invention combines the bright channel prior and the SSR algorithm to preprocess and enhance the images, effectively improving the low visibility of the images and removing noise; the present invention uses threshold-based multiplicative fusion to effectively improve the fusion accuracy of the DI; an unsupervised threshold method based on minimizing the histogram fitting error of the log-normal distribution is proposed, which is applicable to images with very few changed pixels and unchanged scenes under a wide field of view, improving the detection accuracy and robustness, as described in detail below:

[0012] A method for detecting changes in low-light wide-field video images, the method comprising:

[0013] Perform SSR enhancement on the low-light image, and perform principal component analysis fusion on the enhanced image and the bright channel image;

[0014] Based on the enhanced low-light image, use an improved arctangent operator to generate a difference map;

[0015] Combining the chi-square transform and the arctangent operator, an improved multiplicative fusion technique is proposed to process the difference map, and the threshold in the multiplicative fusion technique is processed based on an approximation method of the log-normal distribution to obtain an optimized threshold;

[0016] Process the pixels in the difference map based on minimizing the fitting error of the log-normal distribution, obtain an optimal threshold to segment the difference map, and obtain a change detection result.

[0017] Wherein, the generating of the difference map using the improved arctangent operator is as follows:

[0018]

[0019] Wherein, DI1 represents the arctangent difference map, mse represents the mean square error, which is used to adaptively control the difference degree of different regions, I1(x,y) is the pixel value of the first-phase map at the position (x,y); I2(x,y) is the pixel value of the second-phase map at the position (x,y).

[0020] Among them, the improved multiplication fusion technology processes the difference map as follows:

[0021]

[0022] Among them, T d represents the threshold of image change. DI1 represents the arctangent difference map, DI2 represents the chi-square transform difference map, F represents the multiplication fusion difference map, I1(x, y) is the pixel value of the first temporal phase map at position (x, y); I2(x, y) is the pixel value of the second temporal phase map at position (x, y).

[0023] Among them, the approximation method based on the lognormal distribution processes the threshold in the multiplication fusion technology as follows:

[0024] Let represent the statistical mean and standard deviation corresponding to the two frames of images I1 and I2 respectively, and define the interval as the partial sample sets of the two frames of images I1 and I2, then T d is:

[0025]

[0026] Among them, is the maximum gray level difference between the partial sample sets of the two frames of images.

[0027] Among them, the minimization of the fitting error based on the lognormal distribution processes the pixels in the difference map as follows:

[0028] Let p1(f) and p2(f) represent the probability density functions when the multi-temporal image has changes and no changes respectively. If the multi-temporal image has changes, p1(f) is expressed as:

[0029] p1(f) = p(f|ω u )P(ω u ) + p(f|ω c )P(ω c ), f ∈ (0, 255]

[0030] Among them, f represents the brightness value of a certain pixel in the image, p1(f) represents the PDF when the image has changes, P(ω u ) and P(ω c ) represent the prior probabilities of the unchanged class and the changed class respectively, p(f|ω u ) represents the conditional PDF of the unchanged pixels in DI, p(f|ω c ) represents the conditional PDF of the changed pixels in DI, ω c and ω u represent the changed class and the unchanged class respectively;

[0031] If the multi-temporal images do not change, the probability density function of DI follows a log-normal distribution, and the expression of p2(f) is:

[0032]

[0033] where p2(f) represents the PDF when the image does not change, μ represents the expectation of all pixels, and σ represents the standard deviation of all pixels;

[0034] Define the algorithm optimization model for solving the threshold T as:

[0035]

[0036] where t is a positive number with a relatively small absolute value, and h(f) represents the histogram of DI.

[0037] In a second aspect, a low-light wide-field video image change detection device includes a processor and a memory. Program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method described in any one of the first aspects.

[0038] In a third aspect, a computer-readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, the processor is enabled to execute the method described in any one of the first aspects.

[0039] The beneficial effects of the technical solution provided by the present invention are:

[0040] 1. The present invention adopts a preprocessing enhancement technology that combines the bright channel prior and the single-scale Retinex algorithm to improve the visibility of the image while maintaining its visual naturalness, effectively improving the low visibility of the image and removing noise;

[0041] 2. The present invention proposes multiplicative fusion based on a threshold, which can improve the fusion accuracy of the difference map and enhance the information of the changed area; for the selection of the optimal threshold, since the gray value distribution of the pre-enhanced low-light image is close to a log-normal distribution, this statistical characteristic is used to simplify the computational complexity of threshold calculation;

[0042] 3. The present invention proposes an unsupervised threshold method based on minimizing the histogram fitting error of the log-normal distribution. For images with very few changed pixels and unchanged scenes in low-light wide fields, the threshold segmentation effect obtained is better than that of other global image threshold methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flowchart of a low-light wide-field video image change detection method;

[0044] Figure 2 Histogram schematic diagrams of the pre-enhanced image and the original image;

[0045] Among them, the first row is the phase Figure 1 , and the second row is the phase Figure 2 . (a) is the grayscale image of the original image; (b) is the histogram of the original image; (c) is the grayscale image of the pre-enhanced image; (d) is the histogram of the pre-enhanced image.

[0046] Figure 3 Schematic diagram of the pre-enhancement;

[0047] Among them, (a) is the phase Figure 1 ; (b) is the phase Figure 2 ; (c) is the enhanced phase Figure 1 ; (d) is the enhanced phase Figure 2 .

[0048] Figure 4 Schematic diagram of the change detection result in the low-illumination scene;

[0049] Among them, (a) is the phase Figure 1 ; (b) is the phase Figure 2 ; (c) is the reference image; (d) is the detection result.

[0050] Figure 5 Schematic diagram for detecting the algorithm robustness at different times in the low-illumination scene without change.

[0051] Among them, (a) is the phase Figure 1 ; (b) is the phase Figure 2 ; (c) is the reference image; (d) is the detection result. Specific implementation manners

[0052] To make the objectives, technical solutions and advantages of the present invention clearer, the following further describes the embodiments of the present invention in detail. The overall process is as Figure 1 shown.

[0053] I. Preprocessing - Image enhancement

[0054] Enhancement can improve the visibility of the image while maintaining its visual naturalness. The Retinex-based method has been recognized as a representative technology for this task. The traditional SSR algorithm uses the center-surround theory to estimate the illumination map, resulting in blurred image edges. Therefore, a globally adaptive illumination map is used instead of Gaussian filtering. In addition, since the three channels of the low-illumination image contain different noises, in the embodiments of the present invention, the image is converted from the RGB color space to a grayscale image.

[0055] Thus, the reflection component R(x,y) is obtained, where L w(x,y) represents the luminance value of the low-illumination image, Lg(x,y) represents the brightness value after global adaptation, and σ is a small positive number. The formula is as follows:

[0056]

[0057] Inspired by the dark channel prior defogging method, experimental observations show that the bright channel prior theory is also applicable to low-light images. This theory states that in most regions, at least one color channel has a high brightness. Therefore, the bright channel image under low light can be expressed by the following formula:

[0058]

[0059] In the formula: I c represents an image of a certain color channel of image I; Ω(x) represents the local area centered on pixel point x. After enhancing the low-illumination image with SSR, it is fused with the bright channel image by principal component analysis (PCA) to improve the quality of the enhanced image.

[0060] II. Difference Map Generation and Improved Multiplication Fusion

[0061] Combining the enhanced low-illumination images I1 and I2, the logarithmic operator performs well in dealing with high brightness changes, but has problems in numerical stability and reflecting the true change trend. The mean ratio performs excellently in enhancing the contour of the changed area and preventing information loss, but may introduce noise and affect the retention of details. The extreme pixel ratio operator has advantages in suppressing background information and enhancing information in the changed area, but is too sensitive to changes in small pixel values.

[0062] To overcome this limitation, the embodiment of the present invention uses an improved arctangent operator to generate DI (difference map), which has good dynamic range compression ability. When the pixel difference is large, the arctangent can compress the difference value into a smaller range to avoid over-amplifying the difference. The formula is as follows:

[0063]

[0064] Among them, the mean square error (mse) is used to adaptively control the difference degree of different regions, so that the response in the high-difference value region is not too drastic, while the low-difference value region remains sensitive, effectively retaining local details and suppressing the influence of noise; I1(x, y) is the pixel value at time phase Figure 1 at position (x, y); I2(x, y) is the pixel value at time phase Figure 2 at position (x, y).

[0065] Furthermore, the Chi - Square Transform (CST) effectively distinguishes actual changes from random noise by utilizing its statistical characteristics when generating the DI, providing an accurate measure of change for each pixel by quantifying significant changes in pixel intensity. The robustness of CST reduces the influence of non - changing factors such as illumination and shadows, while the sensitivity to outliers enhances the ability to identify changing objects. Therefore, using CST to generate DI is an effective strategy. The principle of CST is based on the Mahalanobis Distance, and the calculation formula is as follows:

[0066]

[0067] where K is the number of color channels, and DI 1k (x, y) is the difference value of the k - th channel at the pixel (x, y), and μ k is the mean of the difference map of the k - th channel, and σ k is the standard deviation of the difference map of the k - th channel.

[0068] The information of the changed area provided by a single difference map is still affected by residual noise. Although using multi - scale transformation to fuse the DI can improve the accuracy of change detection, it often introduces redundant decomposition, resulting in detail loss and increased computational complexity.

[0069] To further improve the quality of the DI, combining the advantages of CST and the arctangent operator, an improved multiplicative fusion (MTF) technique is proposed, and the formula is as follows:

[0070]

[0071] where T d represents the threshold of possible changes in the image. If T d is set too large, some pixels that may change will not be multiplicatively fused. Conversely, if T d is set too small, some unchanged pixels will be misdetected as changed pixels. In particular, when T d is equal to 0, the improved fusion method is equivalent to MTF.

[0072] Regarding the problem of the optimal threshold T d value, the embodiment of the present invention proposes an approximation method based on the log - normal distribution to obtain the T d value. Since the gray - level distribution of a low - illumination image approximately follows a log - normal distribution after pre - enhancement, using this probability - statistical characteristic can significantly reduce the computational complexity of determining T d .

[0073] Let represent the statistical means and standard deviations corresponding to the two frames of images I1 and I2 respectively. Define the interval is a partial sample set of two frames of images I1 and I2, then T d is:

[0074]

[0075] where is the maximum possible gray - level difference between the partial sample sets of the two frames of images, and 0 is the minimum possible gray - level difference. So T d is approximately the average gray - level difference of these two partial sample sets. In most cases, the interval defined above contains more than half of the pixel number of each frame of the image. Therefore, T d can be approximately used as the threshold for the possible change of the entire image pixels.

[0076] III. Image segmentation by the histogram fitting error minimization method based on log - normal distribution to obtain the change detection map

[0077] When extracting change information from DI, the generally used segmentation methods are the threshold method and the clustering method. The disadvantages of clustering algorithms are that they are sensitive to the selection of initial parameters, noise, and outliers. For example, unsupervised algorithms such as k - means and hierarchical clustering are usually difficult to handle in low - light or noisy environments, resulting in misclassification. This is mainly because these algorithms rely on global statistical information, and under the conditions of low - illumination wide - field, local changes are very small, and the algorithms are prone to failure.

[0078] Therefore, a fitting error minimization method based on log - normal distribution (LNDFEM) is proposed. LNDFEM is derived under the assumptions that the pixel values in DI are independent of each other, the unchanged class approximately follows a log - normal distribution, and the changed class follows a Gaussian distribution. Let p(f|ω u ) represent the conditional PDF of the unchanged pixels in DI, p(f|ω c ) represent the conditional PDF of the changed pixels in DI, ω c and ω u represent the changed class and the unchanged class respectively. Therefore, p(f|ω u ), p(f|ω c ) are defined as follows:

[0079]

[0080] where σ u represents the standard deviation of the unchanged pixels, σ c represents the standard deviation of the changed pixels, μ u represents the expectation of the unchanged pixels, and μ c represents the expectation of the changed pixels.

[0081] Finally, according to the total probability formula, the probability density function of DI can be modeled as composed of ω c and ωu A mixed density distribution composed of two associated density components. There are two cases. The first case is that the multi-temporal images change, and the second case is that there is no change. Therefore, let p1(f) and p2(f) represent the PDFs in these two cases respectively. If the multi-temporal images change, p1(f) can be expressed as:

[0082] p1(f) = p(f|ω u )P(ω u ) + p(f|ω c )P(ω c ), f ∈ (0, 255] (9)

[0083] where P(ω u ) and P(ω c ) represent the prior probabilities of the unchanged class and the changed class respectively.

[0084] If the multi-temporal images do not change, the probability density function of DI follows a log-normal distribution, and the expression of p2(f) is:

[0085]

[0086] where μ represents the expectation of all pixels, and σ represents the standard deviation of all pixels.

[0087] The objective function adopted by this method is to minimize the histogram fitting error, where h(f) represents the histogram of DI.

[0088]

[0089] According to the Bayesian minimum error criterion (maximum a posteriori probability criterion), that is:

[0090]

[0091] To minimize the error probability, each pixel in DI should be assigned to the class with the maximum a posteriori conditional probability, that is:

[0092]

[0093] To find the optimal threshold T, that is, when f = T, p(ω c |f) = p(ω u |f), we get:

[0094]

[0095] If the multi-temporal images change, the fitting effect of p1(f) should be better than that of p2(f). The optimized fitting criterion is defined as follows:

[0096]

[0097] In summary, the objective function is to minimize the fitting error under the conditions of (14) and (15). Therefore, the algorithm optimization model for solving the threshold T is defined as:

[0098]

[0099] where t is a positive number with a relatively small absolute value. The parameters of formulas (7) to (10) are estimated according to the system of equations (17):

[0100]

[0101] IV. Pre-enhanced Histogram Analysis

[0102] From Figure 2 it can be seen that after pre-enhancing the low-light image, its gray-level histogram distribution approximately follows a log-normal distribution. According to this probability and statistical characteristic, the threshold T is calculated d .

[0103] V. Pre-enhanced Quality Analysis

[0104] From Figure 3 it can be seen that after enhancing the low-light image, the contrast of the image is improved, which is convenient for subsequent change detection.

[0105] VI. Low-light Scene Change Detection

[0106] 1) Change detection for low-light scenes with changes

[0107] From Figure 4 (d), it can be seen that the detection results of the embodiments of the present invention in different types of scenes are closer to the reference figure, and at the same time, the edge contours of the changed areas can be better retained, and the false alarm rate is lower.

[0108] 2) Change detection for low-light scenes without changes

[0109] To test the robustness of the invented algorithm, multi-temporal video images of three different scenes are selected for detection. As Figure 5 shown, it can be found from the detection results that there is no influence of redundant noise and no false alarm occurs, indicating that the algorithm has strong robustness.

[0110] A low-light wide-field video image change detection device, the device includes: a processor and a memory, and program instructions are stored in the memory. The processor calls the program instructions stored in the memory to enable the device to execute the following method steps in Embodiment 1:

[0111] Perform SSR enhancement on the low-light image, and perform principal component analysis and fusion on the enhanced image and the bright channel image;

[0112] Based on the enhanced low - illumination image, an improved arctangent operator is used to generate a difference map;

[0113] Combining the chi - square transform and the arctangent operator, an improved multiplication fusion technique is proposed to process the difference map, and based on the approximation method of the log - normal distribution, the threshold in the multiplication fusion technique is processed to obtain an optimized threshold;

[0114] Based on minimizing the fitting error of the log - normal distribution, the pixels in the difference map are processed to obtain an optimal threshold to segment the difference map and get the change detection result.

[0115] Among them, using the improved arctangent operator to generate the difference map is:

[0116]

[0117] Among them, DI1 represents the arctangent difference map, mse represents the mean square error, which is used to adaptively control the difference degree of different regions, I1(x, y) is the pixel value of the first - phase image at the position (x, y); I2(x, y) is the pixel value of the second - phase image at the position (x, y).

[0118] Among them, processing the difference map with the improved multiplication fusion technique is:

[0119]

[0120] Among them, T d represents the threshold of image change. DI1 represents the arctangent difference map, DI2 represents the chi - square transform difference map, F represents the multiplication fusion difference map, I1(x, y) is the pixel value of the first - phase image at the position (x, y); I2(x, y) is the pixel value of the second - phase image at the position (x, y).

[0121] Among them, processing the threshold in the multiplication fusion technique based on the approximation method of the log - normal distribution is:

[0122] Let represent the statistical mean and standard deviation corresponding to the two frames of images I1 and I2 respectively, and define the interval as the partial sample set of the two frames of images I1 and I2, then T d is:

[0123]

[0124] Among them, is the maximum gray - level difference between the partial sample sets of the two frames of images.

[0125] Among them, processing the pixels in the difference map based on minimizing the fitting error of the log - normal distribution is:

[0126] Let \(p_1(f)\) and \(p_2(f)\) denote the probability density functions when there are changes and no changes in the multi-temporal images respectively. If there are changes in the multi-temporal images, \(p_1(f)\) is expressed as:

[0127] \(p_1(f)=p(f|\omega u )P(\omega u ) + p(f|\omega c )P(\omega c ), f\in(0, 255]

[0128] where \(f\) represents the brightness value of a certain pixel in the image, \(p_1(f)\) represents the PDF when there are changes in the image, \(P(\omega u ) and \(P(\omega c ) represent the prior probabilities of the unchanged class and the changed class respectively, \(p(f|\omega u ) represents the conditional PDF of the unchanged pixels in DI, \(p(f|\omega c ) represents the conditional PDF of the changed pixels in DI, \(\omega c and \(\omega u represent the changed class and the unchanged class respectively;

[0129] If there are no changes in the multi-temporal images, the probability density function of DI follows a log-normal distribution, and the expression of \(p_2(f)\) is:

[0130]

[0131] where \(p_2(f)\) represents the PDF when there are no changes in the image, \(\mu\) represents the expectation of all pixels, and \(\sigma\) represents the standard deviation of all pixels;

[0132] Define the algorithm optimization model for solving the threshold \(T\) as:

[0133]

[0134] where \(t\) is a positive number with a relatively small absolute value, and \(h(f)\) represents the histogram of DI.

[0135] It should be noted here that the device description in the above embodiments corresponds to the method description in the embodiments, and the embodiments of the present invention will not be elaborated here.

[0136] The execution subjects of the above-mentioned processor and memory can be devices with computing functions such as computers, single-chip microcomputers, and microcontrollers. Specifically, in implementation, the embodiments of the present invention do not limit the execution subject, and it is selected according to the needs in actual applications. The data signal is transmitted between the memory and the processor through a bus, and the embodiments of the present invention will not elaborate on this.

[0137] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium. The storage medium includes a stored program that controls the device where the storage medium is located to execute the method steps in the above embodiments when the program runs.

[0138] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc. It should be noted here 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 elaborated here.

[0139] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part.

[0140] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through a computer-readable storage medium. The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium or a semiconductor medium, etc.

[0141] Those skilled in the art can understand that the drawings are only schematic diagrams of a preferred embodiment, and the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0142] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting changes in low-illumination wide-field video images, characterized in that, The method includes: Performing SSR enhancement on the low - illumination image and performing principal component analysis fusion on the enhanced image and the bright - channel image; Based on the enhanced low - illumination image, using an improved arctangent operator to generate a difference map; Combining the chi - square transform and the arctangent operator, an improved multiplication fusion technique is proposed to process the difference map, and the threshold in the multiplication fusion technique is processed based on an approximation method of the log - normal distribution to obtain an optimized threshold; Processing the pixels in the difference map based on the minimization of the fitting error of the log - normal distribution, obtaining an optimal threshold - segmented difference map, and getting the change - detection result.

2. The low-illuminance wide-field-of-view video image change detection method according to claim 1, wherein The generating of the difference map using the improved arctangent operator is as follows: Where, DI1 represents the arctangent difference map, mse represents the mean square error, which is used to adaptively control the difference degree of different regions, I1(x,y) is the pixel value of the phase - diagram 1 at the position (x,y); I2(x,y) is the pixel value of the phase - diagram 2 at the position (x,y).

3. A method for detecting changes in low-light wide-field video images according to claim 1, characterized in that, The processing of the difference map by the improved multiplication fusion technique is as follows: Among them, T d represents the threshold of image change, DI1 represents the arctangent difference map, DI2 represents the chi-square transform difference map, F represents the multiplication fusion difference map, I1(x, y) represents the pixel value of the time-phase map 1 at the position (x, y); I2(x, y) represents the pixel value of the time-phase map 2 at the position (x, y).

4. A method for detecting changes in low-light wide-field video images according to claim 1, characterized in that, The processing of the threshold in the multiplication fusion technique based on the approximation method of the log - normal distribution is as follows: Let represent the statistical mean and standard deviation corresponding to two frames of images I1 and I2 respectively, and define the interval as the partial sample sets of two frames of images I1 and I2, then T d is: Among them, is the maximum gray-level difference between two partial sample sets of images.

5. A method for detecting changes in low-light wide-field video images according to claim 1, characterized in that, The processing of the pixels in the difference map based on the minimization of the fitting error of the log - normal distribution is as follows: Let p1(f) and p2(f) represent the probability density functions when the multi - temporal image has changes and no changes respectively. If the multi - temporal image has changes, p1(f) is expressed as: p1(f) = p(f|ω u )P(ω u ) + p(f|ω c )P(ω c ), f ∈ (0, 255] Among them, f represents the brightness value of a certain pixel in the image, p1(f) represents the PDF when the image changes, P(ω u ) and P(ω c ) represent the prior probabilities of the unchanged class and the changed class respectively, p(f|ω u ) represents the conditional PDF of the unchanged pixels in DI, p(f|ω c ) represents the conditional PDF of the changed pixels in DI, ω c and ω u represent the changed class and the unchanged class respectively; If the multi - temporal image has no changes, the probability density function of DI follows the log - normal distribution, and the expression of p2(f) is: Where, p2(f) represents the PDF when the image has no changes, μ represents the expectation of all pixels, and σ represents the standard deviation of all pixels; The algorithm optimization model for solving the threshold T is defined as: Where, t is a positive number with a relatively small absolute value, and h(f) represents the histogram of DI.

6. A low-illuminance wide-field video image change detection device, characterized in that, The device includes: a processor and a memory. Program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method described in any one of claims 1 - 5.

7. A computer-readable storage medium, characterized in that, The computer - readable storage medium stores a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor, the processor executes the method described in any one of claims 1 - 5.

Images