Image data augmentation method and system based on monitoring video

Abstract

The present application relates to the field of image processing, especially to an image data enhancement method and system based on monitoring video, the method comprising: performing connected domain extraction on a target image, calculating connected domain similarity between the target domain and the comparison domain; obtaining a matching connected domain set of the target domain, dividing the connected domain in the target image into dynamic connected domain and static connected domain; obtaining a target scale reflection logarithmic image; calculating the spatiotemporal contrast of the target domain, calculating the fusion weight, and fusing the reflection logarithmic images of all scales in the target channel to obtain a multi-scale fusion image; merging the enhanced images of each channel to obtain the fusion enhanced image of the target image, and completing the image data enhancement. Through the technical scheme of the present application, the accuracy of the image data enhancement result can be improved.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a method and system for image data enhancement based on surveillance video. Background Technology

[0002] Surveillance systems are typically used to collect large amounts of image and video data. This data not only provides rich information for various applications but also offers valuable training material for machine learning and artificial intelligence algorithms. However, images captured in real-world surveillance scenarios are often limited by various factors, such as changes in lighting, weather conditions, shooting angle, and scene congestion. These factors result in inconsistent image quality and significant data loss. Therefore, how to effectively enhance image data from surveillance videos has become a pressing technical challenge.

[0003] In complex monitoring scenarios such as low light, uneven lighting, coexistence of dynamic and static objects, and severe background clutter, existing technologies use the Retinex algorithm to analyze the differences of overall image pixels at different scales, assign weights to different scales, and perform fusion and color restoration to achieve image enhancement. However, this results in the loss of details of dynamic objects in the spatiotemporal dimension, insufficient contrast, artifacts at the edges of brightness changes, and color distortion, leading to inaccurate image enhancement results. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention provides solutions in the following aspects.

[0005] In the first aspect, the image data enhancement method based on surveillance video includes: acquiring a surveillance grayscale image and three channel grayscale images of any sampling frame; using the surveillance grayscale image of any sampling frame as the target image; extracting connected components from the target image; using any connected component as the target domain; using any connected component in the surveillance grayscale image of any sampling frame preceding the target domain as the comparison domain; calculating the connected component similarity between the target domain and the comparison domain; obtaining a set of matching connected components for the target domain based on the connected component similarity; dividing the connected components in the target image into dynamic and static connected components based on the set of matching connected components; and selecting any Gaussian kernel for any channel as the target channel. The width is used as the target scale. Based on the target scale, the grayscale images of the target channels are processed to obtain the logarithmic reflection image at the target scale. The spatiotemporal contrast of the target domain is calculated. Based on the spatiotemporal contrast, the area of ​​the dynamic connected domain, and the area of ​​the static connected domain, the fusion weight of the logarithmic reflection image of the target channel at the target scale is calculated. The logarithmic reflection images of all scales on the target channel are fused to obtain a multi-scale fused image. The product of the multi-scale fused image of the target channel, the color restoration factor of the target channel, and the preset gain coefficient is used as the enhanced image of the target channel. The enhanced images of each channel are merged to obtain the fused enhanced image of the target image, thus completing the image data enhancement.

[0006] Preferably, the calculation of the connected component similarity between the target domain and the comparison domain includes: obtaining the row shape sequence and column shape sequence of the target domain, where each element in the row shape sequence represents the number of pixels in the target domain in the corresponding row, and each element in the column shape sequence represents the number of pixels in the target domain in the corresponding column; constructing a frequency sequence based on the number of times pixel values ​​appear in the target domain; similarly obtaining the row shape sequence, column shape sequence, and frequency sequence of the comparison domain; using the DTW distance between the row shape sequence of the target domain and the row shape sequence of the comparison domain as the first distance, and the DTW distance between the column shape sequence of the target domain and the column shape sequence of the comparison domain as the second distance; calculating the sum of the first distance and the second distance, and using the negative exponent of the sum as the first term; using the exponent of the cosine similarity between the frequency sequence of the target domain and the frequency sequence of the comparison domain as the second term; and using the product of the first term and the second term as the connected component similarity between the target domain and the comparison domain.

[0007] Preferably, the step of obtaining the matching connected component set of the target domain includes: retaining the comparison domains whose connected component similarity is greater than a preset similarity threshold, and obtaining the sampling frame corresponding to any retained comparison domain as a valid frame; for several comparison domains in any valid frame, taking the comparison domain corresponding to the maximum value of the connected component similarity with the target domain as the matching domain of any valid frame, traversing to obtain the matching domain of each valid frame, and constructing all matching domains into the matching connected component set of the target domain.

[0008] Preferably, dividing the connected components in the target graph into dynamic and static connected components includes: calculating the Euclidean distance between the center pixel coordinates of any connected component in the matching connected component set and the center pixel coordinates of the target region, and taking the average of all Euclidean distances as the dynamic distance of the target region; similarly, obtaining the dynamic distance of each connected component in the target graph, calculating the average dynamic distance, and taking connected components whose dynamic distance is greater than the average dynamic distance as dynamic connected components, and taking connected components whose dynamic distance is not greater than the average dynamic distance as static connected components.

[0009] Preferably, the calculation of the spatiotemporal contrast of the target domain includes: constructing a comprehensive domain set by combining all connected components in the matching connected component set of the target domain with the target domain; obtaining the coordinate position of the target domain in the logarithmic reflection image based on the coordinate position of the target domain; calculating the standard deviation of all pixel values ​​in the target domain within the logarithmic reflection image as the pixel standard deviation; similarly obtaining the pixel standard deviation of any connected component in the comprehensive domain set; and using the mean of the pixel standard deviations of each connected component in the comprehensive domain set as the spatiotemporal contrast of the target domain.

[0010] Preferably, the fusion weight includes: taking the sum of the number of pixels in all dynamically connected regions in the target image as the area of ​​the dynamically connected region, and similarly obtaining the area of ​​the static connected regions in the target image; taking the ratio of the area of ​​the dynamically connected region to the area of ​​the static connected region as the dynamic weight, and taking the ratio of the area of ​​the static connected region to the area of ​​the dynamically connected region as the static weight; calculating the spatiotemporal contrast of any dynamically connected region in the target image, and calculating the average of the spatiotemporal contrast of all dynamically connected regions as the first average; calculating the spatiotemporal contrast of any static connected region in the target image, and calculating the average of the spatiotemporal contrast of all static connected regions as the second average; calculating the first product of the dynamic weight and the first average, calculating the second product of the static weight and the second average, and taking the sum of the first product and the second product as the fusion weight.

[0011] Preferably, the fusion weight further includes: for any sampling frame, using the standard deviation of pixel values ​​of all pixels in the log-reflection image corresponding to the target domain as the spatiotemporal contrast of the log-reflection image; for any dynamically connected component in the target image, calculating a first ratio between the spatiotemporal contrast of the dynamically connected component and the spatiotemporal contrast of the log-reflection image, calculating a third product of the first ratio and the dynamic weight, iterating through each dynamically connected component in the target image to obtain the third product, and calculating a first accumulated value of the third product; for any static connected component in the target image, calculating a second ratio between the spatiotemporal contrast of the static connected component and the spatiotemporal contrast of the log-reflection image, calculating a fourth product of the second ratio and the static weight, iterating through each static connected component in the target image to obtain the fourth product, and calculating a second accumulated value of the fourth product; and using the sum of the first accumulated value and the second accumulated value as the fusion weight.

[0012] Secondly, an image data enhancement system based on surveillance video includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the image data enhancement method based on surveillance video described in any one of the claims is implemented.

[0013] The present invention has the following effects:

[0014] This invention effectively improves the accuracy and effect of enhanced surveillance video images by comprehensively utilizing connected component similarity analysis, logarithmic reflection image processing, and adaptive fusion weight calculation, thereby enhancing the identifiability and stability of targets in complex surveillance environments. Specifically, by performing spatiotemporal contrast analysis and shape matching on connected components, dynamic and static regions can be accurately identified, achieving adaptation to the dynamic characteristics of different objects and changes in illumination, thus effectively suppressing the impact of external factors such as sudden changes in illumination and shadow interference on image quality. Furthermore, based on multi-scale extracted logarithmic reflection images, target details can be captured with illumination-invariant features, thus avoiding the problems of lost details in dynamic objects and over-smoothing of static objects in existing technologies, ensuring the accuracy of the enhancement effect. Attached Figure Description

[0015] Figure 1 This is a flowchart of an image data enhancement method based on surveillance video according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0017] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Reference Figure 1 The image data enhancement method based on surveillance video includes steps S1-S5, as detailed below:

[0019] S1: Obtain a monitoring grayscale image and three channel grayscale images of any sampling frame. Use the monitoring grayscale image of any sampling frame as the target image. Extract connected components from the target image. Use any connected component as the target domain. Use any connected component in the monitoring grayscale image of any sampling frame before the target domain as the comparison domain. Calculate the connected component similarity between the target domain and the comparison domain.

[0020] In one embodiment, a monitoring RGB (Red Green Blue) image of any sampling frame is acquired through a monitoring camera, and the monitoring RGB image is converted to grayscale to obtain a monitoring grayscale image. A channel grayscale image is obtained based on the pixel value of each channel in the monitoring RGB image. It should be noted that one monitoring RGB image corresponds to one monitoring grayscale image and three channel grayscale images. The channel refers to the three channels: R, G, and B.

[0021] Randomly select a monitoring grayscale image from the time-series sampling frames as the target image, and perform a connected component extraction algorithm on the target image to obtain several connected components; at the same time, perform the same connected component extraction operation on the monitoring grayscale image of each sampling frame to construct a connected component set corresponding to each frame; any connected component in the target image is used as the target domain; arbitrarily select a connected component from the connected component set of any historical sampling frame before the target image as the comparison domain.

[0022] Obtain the row shape sequence and column shape sequence of the target field. Each element in the row shape sequence represents the number of pixels in the target field in the corresponding row, and each element in the column shape sequence represents the number of pixels in the target field in the corresponding column.

[0023] A frequency sequence is constructed based on the number of times pixel values ​​appear in the target domain; similarly, the row shape sequence, column shape sequence, and frequency sequence of the comparison domain are obtained.

[0024] The first distance is the DTW (Dynamic Time Warping) distance between the row shape sequence of the target domain and the row shape sequence of the comparison domain. The second distance is the DTW distance between the column shape sequence of the target domain and the column shape sequence of the comparison domain. The sum of the first and second distances is calculated, and the negative exponent of the sum is used as the first term. The exponent of the cosine similarity between the frequency sequence of the target domain and the frequency sequence of the comparison domain is used as the second term. The product of the first and second terms is used as the connected component similarity between the target domain and the comparison domain.

[0025] The construction logic of connected component similarity is as follows: the shape distribution of connected components is characterized by row shape sequence and column shape sequence, and the distance between sequences is calculated by DTW algorithm to adapt to different connected component matching; the statistical characteristics of gray-level distribution in connected components are captured by pixel value frequency sequence, and texture consistency is quantified by cosine similarity; the geometric distance is converted into similarity weight by negative exponential function, and multiplied by the exponential value of texture similarity to form a normalized comprehensive similarity index.

[0026] Among them, the elastic matching of shape by DTW distance overcomes the limitations of rigidity measurement, frequency sequence analysis enhances robustness to illumination changes, and the nonlinear transformation and product fusion mechanism of exponential function effectively improve the accuracy of cross-frame target tracking by similarity calculation. In video surveillance scenarios, it can significantly reduce the false matching rate and improve the target recognition ability.

[0027] S2: Obtain the set of matching connected components in the target domain based on the similarity of connected components, and divide the connected components in the target graph into dynamic connected components and static connected components based on the set of matching connected components.

[0028] In one embodiment, a threshold screening is performed on the comparison domains, retaining only those comparison domains whose similarity to the connected components of the target domain exceeds a preset similarity threshold. The sampling frames corresponding to these retained comparison domains are uniformly defined as valid frames (i.e., each valid frame contains at least one high-similarity candidate region). Subsequently, for each valid frame, all connected components (i.e., comparison domains) within the frame are traversed, and the comparison domain with the highest similarity to the connected components of the target domain is selected as the exclusive matching domain for that frame, to ensure that the most relevant region correspondence is captured within a single frame. By traversing all valid frames, the matching domains selected for each valid frame are collected and aggregated to construct the set of matching connected components of the target domain.

[0029] The process of obtaining the set of matching connected components not only effectively filters out low-quality matching noise through the threshold mechanism and improves the robustness of cross-frame association, but also avoids redundancy or conflict through the optimal selection strategy of "one frame, one match", so that the constructed set can accurately reflect the motion trajectory or morphological evolution of the target domain in the time series.

[0030] Calculate the Euclidean distance between the center pixel coordinates of any connected component in the matching connected component set and the center pixel coordinates of the target domain, and take the mean of all Euclidean distances as the dynamic distance of the target domain; similarly, obtain the dynamic distance of each connected component in the target graph, calculate the mean of the dynamic distance, and take the connected components whose dynamic distance is greater than the mean of the dynamic distance as dynamic connected components, and take the connected components whose dynamic distance is not greater than the mean of the dynamic distance as static connected components.

[0031] S3: Take any channel as the target channel, select any Gaussian kernel width as the target scale, and process the channel grayscale image on the target channel based on the target scale to obtain the reflectance logarithmic image of the target scale.

[0032] In one embodiment, a color channel (such as red, green, or blue) is randomly selected from the monitored RGB image as the target channel to focus on specific spectral features; a Gaussian kernel width parameter is specified as the target scale, which controls the spatial smoothing range of the illumination components. The larger the scale, the smoother the captured global illumination changes. An exemplary setting is 15 pixels or 30 pixels; for the monitoring grayscale image extracted from the target frame on the target channel, logarithmic transformation (mapping pixel values ​​to the logarithmic domain, converting multiplicative illumination components into additive forms) and Gaussian smoothing (convolving the logarithmic image with a Gaussian kernel of the selected scale to estimate the smoothed illumination components) are performed sequentially, following the core principle of the Retinex (retinal cortex theory) algorithm: by subtracting the Gaussian smoothed illumination components from the original logarithmic image, the reflection components that only reflect the reflection characteristics of the object surface are separated, and finally the logarithmic reflection image of the target frame at this scale is output.

[0033] For the target channel, the monitoring grayscale image of any historical sampled frame before the target frame is processed in a completely consistent manner at the target scale to generate the corresponding logarithmic reflection image of the historical frame; where one scale corresponds to one logarithmic reflection image.

[0034] S4: Calculate the spatiotemporal contrast of the target domain. Based on the spatiotemporal contrast, the area of ​​the dynamic connected domain, and the area of ​​the static connected domain, calculate the fusion weight of the log reflection image of the target channel at the target scale. Fuse the log reflection images of all scales on the target channel to obtain a multi-scale fused image.

[0035] In one embodiment, the target domain is constructed by combining all connected components in the matching connected component set of the target domain into a comprehensive domain set; the coordinate position of the target domain in the logarithmic reflection image is obtained according to the coordinate position of the target domain, and the standard deviation of all pixel values ​​in the target domain in the logarithmic reflection image is calculated as the pixel standard deviation; similarly, the pixel standard deviation of any connected component in the comprehensive domain set is obtained, and the mean of the pixel standard deviation of each connected component in the comprehensive domain set is used as the spatiotemporal contrast of the target domain.

[0036] The logic for calculating spatiotemporal contrast is as follows: The target domain is aggregated with the set of connected components that match the target domain in historical frames after similarity filtering, forming a comprehensive domain set, which constitutes a complete representation of the target in the spatiotemporal dimension; Based on the coordinate position of the target domain in the original image, its corresponding region in the logarithmic reflection image is accurately located, and the standard deviation of the pixel values ​​in this region is calculated to quantify the dispersion of its internal texture details; Similarly, this operation is repeated for each connected component in the comprehensive domain set to obtain its respective pixel standard deviation; Finally, the mean of these standard deviations is defined as the spatiotemporal contrast of the target domain.

[0037] Spatiotemporal contrast integrates spatial texture variations (pixel fluctuations within a single frame) with temporal dynamic evolution (regional consistency across multiple frames), thereby effectively suppressing the effects of sudden illumination changes, shadow interference, and short-term noise, and significantly improving the robustness of target representation.

[0038] The fusion weights of the target channel's reflectance logarithmic image at the target scale are calculated based on spatiotemporal contrast, the area of ​​the dynamically connected region, and the area of ​​the statically connected region.

[0039] It should be noted that in the field of surveillance video image enhancement, existing technologies typically use multi-scale Gaussian kernels to process each frame of the image, allocating the enhancement contribution of each scale through global pixel differences or equal weights. However, this method fails to distinguish the essential difference in spatiotemporal contrast stability between dynamic and static objects: dynamic regions experience drastic contrast fluctuations due to motion, while static regions are relatively stable. As a result, in low-light or uneven lighting scenarios, details of dynamic objects are easily lost through smoothing, and edges with sudden brightness changes are prone to artifacts and color distortion. This invention generates logarithmic reflection images at different scales based on the Retinex algorithm to separate illumination-invariant features; for each scale, it accurately identifies and quantifies the contrast stability and relative area ratio of dynamically connected and statically connected regions, adaptively calculating the fusion weights for each scale, so that high weights are allocated to scales that can simultaneously preserve dynamic domain details (such as high texture variation) and static domain smoothness (such as low-noise edges).

[0040] The sum of the number of pixels in all dynamically connected regions in the target image is used as the area of ​​the dynamically connected region. Similarly, the area of ​​the static connected regions in the target image is obtained. The ratio of the area of ​​the dynamically connected region to the area of ​​the static connected region is used as the dynamic weight, and the ratio of the area of ​​the static connected region to the area of ​​the dynamically connected region is used as the static weight. The spatiotemporal contrast of any dynamically connected region in the target image is calculated, and the mean of the spatiotemporal contrast of all dynamically connected regions is calculated as the first mean. The spatiotemporal contrast of any static connected region in the target image is calculated, and the mean of the spatiotemporal contrast of all static connected regions is calculated as the second mean. The first product of the dynamic weight and the first mean is calculated, and the second product of the static weight and the second mean is calculated. The sum of the first product and the second product is used as the fusion weight.

[0041] A multi-scale fused image is obtained by fusing the logarithmic reflection images of all scales on the target channel.

[0042] In another embodiment, the fusion weight further includes: for any sampling frame, using the standard deviation of pixel values ​​of all pixels in the log-reflection image corresponding to the target domain as the spatiotemporal contrast of the log-reflection image.

[0043] For any dynamically connected component in the target graph, calculate the first ratio of the spatiotemporal contrast of the dynamically connected component to the spatiotemporal contrast of the logarithmic reflection image, calculate the third product of the first ratio and the dynamic weights, iterate through each dynamically connected component in the target graph to obtain the third product, and calculate the first accumulated value of the third product; for any static connected component in the target graph, calculate the second ratio of the spatiotemporal contrast of the dynamically connected component to the spatiotemporal contrast of the logarithmic reflection image, calculate the fourth product of the second ratio and the static weights, iterate through each static connected component in the target graph to obtain the fourth product, and calculate the second accumulated value of the fourth product; use the sum of the first accumulated value and the second accumulated value as the fusion weight.

[0044] It should be noted that area analysis quantifies the pixel coverage ratio of connected components in a single frame image. Objects with larger areas need to receive higher adaptive weights during enhancement to ensure that the multi-scale fusion algorithm prioritizes their structural integrity and detail capture capabilities, avoiding excessive smoothing or loss of detail due to global equal enhancement. Meanwhile, spatiotemporal contrast dynamically evaluates the contrast stability of objects in the time dimension by calculating the mean of the standard deviation of pixels of the same connected component in consecutive frame logarithmic reflection images (e.g., high values ​​correspond to fast-moving dynamic targets, while low values ​​correspond to static areas with stable lighting), thereby distinguishing the motion attributes of objects from environmental interference.

[0045] S5: The product of the multi-scale fused image of the target channel, the color restoration factor of the target channel, and the preset gain coefficient is used as the enhanced image of the target channel. The enhanced images of each channel are merged to obtain the fused enhanced image of the target image, thus completing the image data enhancement.

[0046] In one embodiment, a multi-scale fused image of the target channel is generated; a color restoration factor (dynamically generated by calculating the ratio of the target channel to the mean of all channels in the original image to compensate for the color shift caused by Retinex processing) is introduced and multiplied by a preset gain coefficient (used to adjust the overall brightness dynamic range, usually adaptively set according to the scene illumination, for example, 1.5-2.0 in low-light scenes) to perform color correction and contrast expansion on the fused image; finally, the calibrated target channel enhanced image is recombined with the remaining channels (processed in the same way) to generate a fused enhanced image of the target image through linear superposition or nonlinear mapping.

[0047] It should be noted that the method for obtaining the color restoration factor is a well-known technique to those skilled in the art, and will not be elaborated upon here.

[0048] The system includes a processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the image data enhancement method based on surveillance video according to the first aspect of the present invention.

[0049] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0050] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An image data enhancement method based on surveillance video, characterized in that, include: Obtain a monitoring grayscale image and three channel grayscale images of any sampling frame. Use the monitoring grayscale image of any sampling frame as the target image. Extract connected components from the target image. Use any connected component as the target domain. Use any connected component in the monitoring grayscale image of any sampling frame before the target domain as the comparison domain. Calculate the connected component similarity between the target domain and the comparison domain. The target domain is obtained by matching the set of connected components based on the similarity of connected components. Based on the set of matching connected components, the connected components in the target graph are divided into dynamic connected components and static connected components. Take any channel as the target channel, select any Gaussian kernel width as the target scale, and process the channel grayscale image on the target channel based on the target scale to obtain the reflectance logarithmic image of the target scale. Calculate the spatiotemporal contrast of the target domain, and calculate the fusion weight of the log reflection image of the target channel at the target scale based on the spatiotemporal contrast, the area of ​​the dynamic connected domain, and the area of ​​the static connected domain. Then, fuse the log reflection images of all scales on the target channel to obtain a multi-scale fused image. The multi-scale fused image of the target channel, the color restoration factor of the target channel, and the preset gain coefficient are multiplied to obtain the enhanced image of the target channel. The enhanced images of each channel are then merged to obtain the fused enhanced image of the target image, thus completing the image data enhancement.

2. The image data enhancement method based on surveillance video according to claim 1, characterized in that, The calculation of the connected component similarity between the target domain and the comparison domain includes: Obtain the row shape sequence and column shape sequence of the target field. Each element in the row shape sequence represents the number of pixels in the target field in the corresponding row, and each element in the column shape sequence represents the number of pixels in the target field in the corresponding column. Construct a frequency sequence based on the number of times pixel values ​​appear within the target domain; Similarly, the row shape sequence, column shape sequence, and frequency sequence of the alignment field are obtained; The DTW distance between the row shape sequence of the target domain and the row shape sequence of the alignment domain is taken as the first distance, and the DTW distance between the column shape sequence of the target domain and the column shape sequence of the alignment domain is taken as the second distance. The sum of the first distance and the second distance is calculated, and the negative exponent of the sum is taken as the first term. The exponential value of the cosine similarity between the frequency sequence of the target domain and the frequency sequence of the comparison domain is used as the second term; The product of the first and second terms is used as the similarity of the connected components between the target domain and the comparison domain.

3. The image data enhancement method based on surveillance video according to claim 1, characterized in that, The set of matching connected components for the target domain includes: The comparison domains whose connected component similarity is greater than a preset similarity threshold are retained, and the sampled frame corresponding to any retained comparison domain is taken as a valid frame. For any valid frame, the matching domain corresponding to the maximum similarity of the connected components with the target domain is taken as the matching domain of any valid frame. The matching domain of each valid frame is obtained by traversing through the matching domains, and all matching domains are constructed as the set of matching connected components of the target domain.

4. The image data enhancement method based on surveillance video according to claim 1, characterized in that, The step of dividing the connected components in the target graph into dynamic connected components and static connected components includes: Calculate the Euclidean distance between the center pixel coordinates of any connected component in the matching connected component set and the center pixel coordinates of the target domain, and use the mean of all Euclidean distances as the dynamic distance of the target domain; Similarly, obtain the dynamic distance of each connected component in the target graph, calculate the mean dynamic distance, and designate connected components whose dynamic distance is greater than the mean dynamic distance as dynamic connected components, and those whose dynamic distance is not greater than the mean dynamic distance as static connected components.

5. The image data enhancement method based on surveillance video according to claim 1, characterized in that, The spatiotemporal contrast of the target domain for calculation includes: Construct a comprehensive domain set by combining all connected components in the set of connected components that match the target domain with the target domain; The coordinates of the target domain in the logarithmic reflection image are obtained based on the coordinates of the target domain. The standard deviation of all pixel values ​​in the target domain within the logarithmic reflection image is calculated as the pixel standard deviation. Similarly, the pixel standard deviation of any connected domain in the comprehensive domain set is obtained, and the mean of the pixel standard deviation of each connected domain in the comprehensive domain set is used as the spatiotemporal contrast of the target domain.

6. The image data enhancement method based on surveillance video according to claim 1, characterized in that, The fusion weights include: The area of ​​a dynamic connected region is the sum of the number of pixels in all dynamically connected regions in the target graph. Similarly, the area of ​​a static connected region in the target graph can be obtained. The ratio of the area of ​​the dynamically connected component to the area of ​​the statically connected component is used as the dynamic weight, and the ratio of the area of ​​the statically connected component to the area of ​​the dynamically connected component is used as the static weight. Calculate the spatiotemporal contrast of any dynamic connected component in the target graph, and calculate the mean of the spatiotemporal contrast of all dynamic connected components as the first mean; calculate the spatiotemporal contrast of any static connected component in the target graph, and calculate the mean of the spatiotemporal contrast of all static connected components as the second mean. Calculate the first product of the dynamic weight and the first mean, calculate the second product of the static weight and the second mean, and use the sum of the first product and the second product as the fusion weight.

7. The image data enhancement method based on surveillance video according to claim 6, characterized in that, The fusion weights also include: For any sampling frame, the standard deviation of the pixel values ​​of all pixels in the log-reflection image corresponding to the target domain is used as the spatiotemporal contrast of the log-reflection image; For any dynamically connected component in the target graph, calculate the first ratio of the spatiotemporal contrast of the dynamically connected component to the spatiotemporal contrast of the logarithmic reflection image, calculate the third product of the first ratio and the dynamic weight, traverse to obtain the third product of each dynamically connected component in the target graph, and calculate the first accumulated value of the third product. For any static connected component in the target graph, calculate the second ratio of the spatiotemporal contrast of the static connected component to the spatiotemporal contrast of the logarithmic reflection image, calculate the fourth product of the second ratio and the static weight, traverse to obtain the fourth product of each static connected component in the target graph, and calculate the second accumulated value of the fourth product. The sum of the first and second accumulated values ​​is used as the fusion weight.

8. An image data enhancement system based on surveillance video, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the image data enhancement method based on surveillance video according to any one of claims 1-7.

Images