A method and apparatus for processing surveillance playback images based on conditional generative adversarial networks

Abstract

This invention relates to the field of surveillance video processing technology. It provides a method and apparatus for processing surveillance playback footage based on conditional generative adversarial networks (GANs). The method includes: collecting degraded video frame samples under multiple degradation conditions and corresponding reference frame samples based on the surveillance service scenario, establishing a training dataset with degradation types and clarity pairings; constructing a video processing model based on a GAN, training the video processing model using the constructed training dataset to obtain a trained video processing model; inputting the video to be processed into the trained video processing model to obtain enhanced and optimized video data; and performing an index query through a retrieval mechanism to obtain the playback video corresponding to the user's request. This invention effectively improves the quality of surveillance playback footage.

Description

Technical Field

[0001] This invention belongs to the field of video processing technology, and provides a method and apparatus for processing surveillance playback images based on conditional generative adversarial networks. Background Technology

[0002] With the rapid development of intelligent surveillance technology, surveillance video is increasingly widely used in public safety, traffic management, smart homes, and other fields. However, in practical applications, the quality of surveillance video is often constrained by various environmental factors, such as low light, backlight, low resolution, and blurry images. These factors severely affect the clarity and usability of the video, thereby reducing the effectiveness and reliability of the surveillance system. Traditional image enhancement methods, such as histogram equalization and filtering, can improve image quality in specific scenarios, but their effects are limited and they are difficult to adapt to complex and changing surveillance environments. In recent years, Generative Adversarial Networks (GANs), as a powerful generative model, have shown great potential in the field of image enhancement. In particular, Conditional Generative Adversarial Networks (CGANs), by introducing conditional information (such as lighting and resolution), can generate high-quality images for specific scenarios, providing a new solution for enhancing surveillance video.

[0003] Histogram equalization and filtering are two classic image enhancement methods widely used to improve the quality of surveillance playback footage. Histogram equalization redistributes the gray values ​​of image pixels, making the gray-level distribution more uniform, thus significantly improving image contrast and detail, especially suitable for surveillance footage in low light or with insufficient contrast. Filtering, on the other hand, uses specific filters (such as Gaussian filtering and median filtering) to remove noise or blur from the image while preserving important edge and detail information. While both methods effectively improve image quality to some extent, histogram equalization can lead to over-enhancement in local areas, causing loss of detail or artifacts; filtering, while removing noise, can blur image edges and details, affecting image clarity. Images generated by histogram equalization and filtering can sometimes appear unnatural and prone to showing artificial artifacts.

[0004] Furthermore, while traditional histogram equalization methods can improve overall contrast in low-light or low-contrast surveillance footage, they can easily lead to over-enhancement in local areas, causing loss of detail or artifacts, affecting the visual quality and accuracy of subsequent analysis. Existing filtering methods, while removing image noise, can blur image edges and details, even reducing image clarity, which is particularly detrimental to surveillance scenarios requiring high-precision target recognition and behavior analysis. Traditional image enhancement methods often produce unnatural-looking images when processing complex surveillance footage, easily revealing artificial artifacts and reducing visual quality and realism. Current technologies struggle to dynamically adjust enhancement strategies in complex and ever-changing surveillance environments, and cannot effectively handle complex scenes with multiple degradation conditions (such as low light, backlight, low resolution, and motion blur) simultaneously, resulting in limited enhancement effects.

[0005] In summary, while histogram equalization and filtering perform well in simple scenarios, they have significant shortcomings when processing complex monitoring footage. In other words, there is still considerable room for improvement in optimizing monitoring footage.

[0006] Therefore, it is necessary to provide a monitoring playback processing method and apparatus based on conditional generative adversarial networks to solve the above problems. Summary of the Invention

[0007] This invention provides a monitoring playback image processing method and apparatus based on conditional generative adversarial networks to solve the problems of traditional histogram equalization methods in the prior art when processing monitoring images with low light or insufficient contrast. These problems include excessive enhancement of local areas due to increasing overall contrast, loss of details or artifacts, blurring of image edges and details due to noise removal, and even reduced image clarity. Furthermore, these methods struggle to adapt to dynamic adjustments in complex and changing monitoring environments, cannot effectively handle various degradation conditions, and have limited enhancement effects. The invention addresses these technical problems by implementing the following technical solutions.

[0008] The first aspect of this invention proposes a monitoring playback image processing method based on a conditional generative adversarial network (GAN). The method includes: collecting degraded video frame samples under multiple degradation conditions and corresponding reference frame samples based on a monitoring service scenario; establishing a training dataset with degradation type and clarity pairings; the degraded video frame samples include video frame samples of the following degradation types: low-light type, backlight type, low-resolution type, and motion blur type; constructing a video processing model based on a GAN; and training the video processing model using the constructed training dataset to obtain a trained video processing model. Specifically, during the training process, the degradation type is mapped to a specified dimension. Conditional vectors are fused with the video to be processed, which is the input to the model, to generate degradation-aware features. Local texture details are enhanced and noise is suppressed through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules. Hierarchical optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss. The video to be processed is input into the trained video processing model to obtain enhanced and optimized video data. The obtained enhanced and optimized video data is spatiotemporally sliced ​​to form structured information storage. A bidirectional hash mapping table between slice storage addresses and time intervals is established. An index query process is executed through a retrieval mechanism to obtain the playback video corresponding to the user's request.

[0009] The second aspect of this invention proposes a monitoring playback image processing device based on a conditional generative adversarial network (GAN), which is the monitoring playback image processing method based on a conditional GAN ​​described in the first aspect of this invention. The monitoring playback image processing device includes: an acquisition and processing module, which, based on a monitoring business scenario, acquires degraded video frame samples under multiple degradation conditions and reference frame samples corresponding to the business scenario, and establishes a training dataset with degradation type and clarity pairing. The degraded video frame samples include video frame samples of the following degradation types: low-light type, backlight type, low-resolution type, and motion blur type; and a model building module, which, based on a conditional GAN, constructs a video processing model and trains the video processing model using the constructed training dataset to obtain a trained video processing model, specifically including: in During training, degradation types are mapped to conditional vectors of specified dimensions and fused with the video to be processed as model input to generate degradation-aware features. Deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules enhance local texture details and suppress noise. Layered optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss. The output module inputs the video to be processed into the trained video processing model to obtain enhanced and optimized video data. The query and retrieval module spatiotemporally slices the obtained enhanced and optimized video data to form structured information storage, establishes a bidirectional hash mapping table between slice storage addresses and time intervals, and executes an index query process through a retrieval mechanism to obtain the playback video corresponding to the user's request.

[0010] A third aspect of the present invention provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the monitoring playback image processing method based on conditional generative adversarial networks as described in the first aspect of the present invention.

[0011] A fourth aspect of the present invention provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the monitoring playback processing method based on conditional generative adversarial networks as described in the first aspect of the present invention.

[0012] The embodiments of the present invention have the following advantages: Compared with existing technologies, this invention constructs a video processing model based on conditional generative adversarial networks. The video processing model is trained using the constructed training dataset to obtain a well-trained model. Inputting the video to be processed into the trained model yields enhanced and optimized video data, effectively improving the visual effect of the surveillance playback images. The enhanced and optimized video data is then spatiotemporally sliced ​​to form structured information storage. A bidirectional hash mapping table between the sliced ​​storage addresses and time intervals is established, and an index query is performed through a retrieval mechanism to improve query efficiency.

[0013] Specifically, by introducing a conditional generative adversarial network (GAN), the generation strategy is dynamically adjusted to avoid over-enhancement in local areas while preserving and enhancing image details. This effectively improves the visual quality of surveillance playback images and the accuracy of subsequent analysis. A hybrid domain generator is used, combining spatial and frequency domain dual-path processing mechanisms. Through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules, local texture details are enhanced while noise is suppressed, ensuring that edges and details in surveillance playback images are clearly visible. Through a multi-granularity discriminator and an adaptive loss function system, more natural and higher-quality enhanced images of surveillance playback images are generated, reducing human intervention and improving the visual quality and realism of surveillance playback images.

[0014] Furthermore, the dynamic conditional coding module based on conditional generative adversarial networks can adapt to complex and ever-changing monitoring environments, effectively handle various degradation scenarios, and significantly improve the clarity and detail of monitoring playback videos. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an example of the monitoring playback processing method based on conditional generative adversarial networks of the present invention. Figure 2 This is a flowchart illustrating another aspect of the monitoring playback image processing method based on conditional generative adversarial networks of the present invention. Figure 3 This is a schematic diagram of the architecture of the conditional generative adversarial network in this invention; Figure 4 This is a structural block diagram of the monitoring playback image processing device based on conditional generative adversarial networks of the present invention. Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a computer-readable medium embodiment according to the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] In view of the above problems, this invention proposes a method and apparatus for processing surveillance playback images based on conditional generative adversarial networks (GANs). This method, based on a surveillance business scenario, collects degraded video frame samples under multiple degradation conditions and corresponding reference frame samples for the business scenario, establishing a training dataset with degradation type and clarity pairings. The degraded video frame samples include video frame samples of the following degradation types: low-light type, backlight type, low-resolution type, and motion blur type. Based on a conditional generative adversarial network, a video processing model is constructed, and the constructed training dataset is used to train the video processing model to obtain a trained video processing model. Specifically, during the training process, the degradation type is mapped to a reference frame. A fixed-dimensional conditional vector is fused with the video to be processed, which serves as the model input, to generate degradation-aware features. Local texture details are enhanced and noise is suppressed through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules. Hierarchical optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss. The video to be processed is input into the trained video processing model to obtain enhanced and optimized video data. The obtained enhanced and optimized video data is spatiotemporally sliced ​​to form structured information storage. A bidirectional hash mapping table between slice storage addresses and time intervals is established, and an index query process is executed through a retrieval mechanism to improve query efficiency.

[0018] Example 1 The following reference Figure 1 , Figure 2 , Figure 3 The present invention will be described in detail below.

[0019] Figure 1 This is a flowchart illustrating an example of the monitoring playback processing method based on conditional generative adversarial networks of the present invention. Figure 2 This is a flowchart illustrating another aspect of the monitoring playback processing method based on conditional generative adversarial networks of the present invention.

[0020] like Figure 1 As shown, in step S101, based on the monitoring business scenario, degraded video frame samples under multiple degradation conditions and reference frame samples of the corresponding business scenario are collected to establish a training dataset with degradation type and clarity pairing. The degraded video frame samples include video frame samples with the following degradation types: low light type, backlight type, low resolution type and dynamic blur type.

[0021] Specifically, surveillance scenarios include: traffic monitoring (such as roads, intersections, highways, and tunnels), which involves issues such as blurred vehicle movement, low light (nighttime), or backlight (strong light); security monitoring (such as residential areas, shopping malls, banks, and public places), which faces degradation due to low resolution (long-distance shooting) or blurred movement (fast-moving targets); industrial monitoring (such as factory production lines and warehousing logistics), which suffers from reduced video quality due to insufficient ambient light or equipment vibration; urban public monitoring (such as large public places like squares, subway stations, and airports), which suffers from video degradation due to complex lighting conditions or dense crowds; and special environment monitoring (such as forest fire prevention and border patrol), which suffers from video quality issues due to weather conditions (fog, rain, snow) or long-distance shooting.

[0022] By synchronously acquiring specified quality reference frame samples for the corresponding business scenario using professional image acquisition equipment, a training dataset with degradation type and sharpness pairing (i.e., degradation-sharpness pairing) is established, and standardized preprocessing and data augmentation operations are performed.

[0023] The training dataset is created using the following expression: ; in, The training dataset representing the degradation type and the clear pairing (i.e., the degradation type and the reference frame pairing); Let N be a sample of degraded video frames of the i-th degradation type, where N represents the number of degradation types, and i and N are both positive integers, i being 1, 2, 3, ..., N; This is the reference frame sample (e.g., 1920×1080) corresponding to the degraded video frame sample of the i-th degradation type, where H represents the vertical number of pixels in the video frame and W represents the horizontal number of pixels in the video frame. and of , ; This is the label for the i-th degradation type.

[0024] Through normalization preprocessing (e.g., ImageNet normalization): the input degraded video frame samples and reference frame samples Channel-by-channel mean subtraction and variance scaling are performed separately. A hybrid enhancement strategy is also executed, specifically including random cropping: from, for example, degraded video frame samples... and reference frame samples The process involves several key techniques: randomly cropping sub-regions of a specific size (e.g., 256×256) from the original frames to ensure they contain key content; color dithering: randomly perturbing brightness, contrast, saturation, and hue; multi-degradation linear fusion: linearly fusing video frame samples of different degradation types with random weights to generate composite degradation samples; random horizontal flipping: horizontally flipping frames with a 50% probability; Gaussian noise injection: adding zero-mean Gaussian noise to degraded frames; and dynamic blur simulation: applying blur to clear frames using randomly generated motion kernels to simulate dynamic blur degradation. This expands data diversity, providing a learning foundation for accurate degradation perception and enhancement capabilities in conditional generative adversarial networks (GANs).

[0025] It should be noted that, in this invention, the business scenario characteristics specifically include the selection of the specified duration, which is determined based on specific business needs and the characteristics of the monitoring scenario. The selection of the specified duration considers the following factors: event frequency (if the event frequency in the monitoring scenario is high, a shorter sharding duration, 5 to 45 seconds, is configured to capture more event details), storage and computing resource limitations (shorter sharding durations increase storage and computing overhead, therefore, an appropriate duration is selected within the resource allowance), and the granularity of business analysis (the sharding duration matches the granularity of business analysis; for example, if minute-level behavior needs to be analyzed, the sharding duration can be configured at the minute level, such as 5 to 10 minutes). The above are merely optional examples and should not be construed as limiting the invention.

[0026] The choice of video segment duration needs to be dynamically adjusted based on the characteristics and requirements of different monitoring scenarios. In traffic monitoring scenarios, such as roads and intersections, short segments of 5 to 10 seconds are used to capture high-dynamic events; while for highways and tunnels, the duration is appropriately extended to 15 to 30 seconds to balance storage and retrieval efficiency. In security monitoring, critical areas such as residential communities and banks use segments of 10 to 30 seconds, while shopping malls and public places are set to 1 to 2 minutes to balance event analysis and storage optimization. Industrial monitoring scenarios (such as production lines) use segments of 30 seconds to 1 minute to match production rhythm; warehousing and logistics are relaxed to 1 to 2 minutes. Urban public monitoring (such as subway stations and squares) uses segments of 1 to 5 minutes due to dense crowds and sparse events. Special environment monitoring (such as forest fire prevention) uses segments of 5 to 10 minutes to reduce redundant data due to rare events. By flexibly configuring the segment duration, the use of storage and computing resources can be optimized while meeting business analysis needs.

[0027] It should be noted that the above is only an optional example and should not be construed as a limitation of the present invention.

[0028] Next, in step S102, a video processing model is constructed based on a conditional generative adversarial network. The video processing model is trained using the constructed training dataset to obtain a trained video processing model. Specifically, during the training process, degradation types are mapped to conditional vectors of a specified dimension and fused with the video to be processed as the model input to generate degradation-aware features. Local texture details are enhanced and noise is suppressed through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules. Hierarchical optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss.

[0029] Specifically, a conditional generative adversarial network (GAN) is used to construct the video processing model. The GAN consists of a generator and a discriminator. The generator generates data, and the discriminator distinguishes between the generated data (in this example, "pseudo-high-quality video frames" generated by the generator based on input degraded video frame samples (such as low-light, backlight, low-resolution, or motion-blurred video frames) and their corresponding degradation type conditional information (such as low-light labels) and the real data (in this example, reference frame samples in the training dataset, i.e., high-quality video frames paired with degraded video frame samples). The performance of both is improved through adversarial training. The conditional generative adversarial network (CGAN) introduces conditional information on top of the GAN, enabling the generator to generate high-quality data based on these conditions. In this task, the generator receives low-quality degraded video frame samples (such as low-light, backlight, low-resolution, or motion-blurred video frames) and their corresponding degradation type conditional information (such as low-light labels), and generates high-quality video frames using structures such as convolutional neural networks (CNNs). The goal is to make the generated video frames as close as possible to the specified quality reference frame samples (HD 1920×1080 resolution) for the corresponding business scenario in terms of visual quality (PSNR ≥ 30dB, SSIM ≥ 0.9) and detail reproduction. The discriminator receives the generated high-quality video frames and real reference frame samples, and judges their authenticity based on multi-granularity (global, local, temporal) analysis. The aim is to accurately distinguish between generated frames and real reference frames, and through adversarial training and co-optimization with the generator, ultimately improve the model's ability to enhance videos with multiple degradation types.

[0030] The conditional generative adversarial network architecture maps degenerate types to high-dimensional conditional vectors through a dynamic conditional coding module. Specifically, the dynamic conditional coding module maps discrete degenerate type labels to continuous conditional vectors of a specified dimension. First, the discrete degenerate type labels... Perform One-Hot encoding to generate an initial vector. (There are 4 types of degradation). Next, the initial vector is processed through a fully connected layer (FC) and the LeakyReLU activation function. A conditional vector mapped to a specified dimension (100 to 300 dimensions, such as 128 or 132 dimensions). :

[0031] ; in, The weight matrix is ​​a learnable matrix, optimized using the backpropagation gradient descent algorithm. For bias terms, It encodes global semantic information about the degenerate type; Indicates the label for discrete degeneracy type The initial vector generated by One-Hot encoding.

[0032] Input the video frame to be processed Spatial features are extracted using pre-trained shallow CNNs (e.g., 3-layer convolutions). (Specifically including local structures and textures related to degradation), the shallow CNN structure specifically includes: Convolutional layer 1: Convolution, 32 output channels, stride 2, ReLU; Convolutional layer 2: Convolution, 64 output channels, stride 2, ReLU; Convolutional layer 3: Convolution, output 64 channels, stride 1. The spatial features include low-level texture features, local structural features, and channel dimension representations associated with the four degradation types.

[0033] The specified dimension conditional vector representing the degradation type is fused with the features of the input video frame. Adjusted via broadcast mechanism To maintain spatial consistency, the dimensions are then stitched together along the channel dimension. and the expanded ,get Finally passed Convolutional dimensionality reduction to generate degenerate perception features It integrates degradation types with spatial characteristics.

[0034] For the hybrid domain generator, a U-Net++ backbone network is employed, combining spatial and frequency domain dual-path processing mechanisms. Through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules, local texture details (specifically, in this example, low-level texture features, local structural features, and high-frequency details related to the four degradation types) are enhanced, and noise is suppressed. Finally, the enhanced result is output through a gated fusion unit. For details, please refer to [link to documentation]. Figure 3 .

[0035] Dense skip connections are introduced on top of the standard U-Net to enhance multi-scale feature fusion.

[0036] Encoder No. Layer output: ; in: Indicates the encoder's first... The output feature map of the layer; This represents a 3×3 convolution operation; This represents the modified linear unit activation function.

[0037] Decoder Layer input: ; in, Indicates the decoder's first The input feature map of the layer; Indicates the first The input feature map of the layer is upsampled twice bilinearly; d represents the decoder's first... The output feature map of the layer; e represents the output feature map of the encoder; This indicates that a feature map splicing operation is being performed.

[0038] The spatial domain path of the dual-path processing mechanism employs deformable convolution to calculate the offset of each sampling point relative to the regular grid position in the deformable convolution: ; in: This represents the offset of each sampling point in deformable convolution relative to the position of the regular grid, enhancing local details (such as repairing blurred edges). This represents the input feature map to the deformable convolutional module.

[0039] The output features of a deformable convolutional module can be represented by the following expression: ; in, This represents the output feature map of the deformable convolutional module, where p represents the spatial coordinates on the feature map. For regular grid positions; Indicates the convolution kernel weights, The kernel size is indicated (e.g., 9), a mechanism that allows the network to adaptively adjust its receptive field.

[0040] In the channel attention module, the channel attention weights are represented as follows: ; in, Indicates channel attention weights; Indicates global average pooling; Indicates a fully connected layer; This represents the Sigmoid activation function; ,Right now This represents the output feature map of the deformable convolutional module.

[0041] Output characteristics of the channel attention module: ; in, This represents the output characteristics of the channel attention module; This indicates the channel attention weight, highlighting important channels; This indicates a channel-by-channel multiplication operation.

[0042] The frequency domain path of the dual-path processing mechanism uses Fast Fourier Transform (FFT) to represent the frequency domain: ; in, This represents the frequency domain representation of the input feature map after Fast Fourier Transform, where u and v represent the horizontal and vertical coordinates of the input feature map in the frequency domain, respectively. These represent the width and height of the feature map input to the frequency domain path, respectively; Represents the imaginary unit; (x, y) represents the position information of the input feature map to the deformable convolutional module in the frequency domain path, where x represents the horizontal coordinate in the frequency domain path and y represents the vertical coordinate in the frequency domain path.

[0043] Frequency domain representation Perform frequency domain masking: ; in, The mask weights represent the frequency domain weights, where u and v represent the horizontal and vertical coordinates of the input feature map in the frequency domain, respectively; the mask M suppresses noise frequency bands. This indicates the amplitude of the spectrum.

[0044] The filtered spectrum is represented as follows: ; in, This represents the filtered spectrum; The mask weights in the frequency domain are represented by u and v, which represent the horizontal and vertical coordinates of the input feature map in the frequency domain, respectively. This represents the spectrum before filtering.

[0045] Finally, the enhanced result is output through the gated fusion unit: ; in, The fusion weights represent the output features of the spatial domain path and the output features of the frequency domain path; This represents the output characteristics of the spatial domain path; This represents the output characteristics of the frequency domain path.

[0046] The final output is represented as: ; in, represents the final output of the dual-path processing mechanism involving both spatial and frequency domain paths; g represents the weights generated by the gated fusion unit. This represents the output characteristics of the spatial domain path; This represents the output characteristics of the frequency domain path; The fusion weight represents the output features of the spatial domain path and the output features of the frequency domain path.

[0047] The multi-granularity discriminator includes global, local, and temporal discriminators, and WGAN-GP is used to improve training stability. The adaptive loss function system combines degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss to achieve hierarchical optimization.

[0048] Specifically, the global discriminator is a five-layer CNN+FC, with a scope covering the entire frame of the image; the local discriminator is a four-layer CNN+FC, with a scope covering randomly cropped patches; and the temporal discriminator uses 3DCNN+LSTM, with a scope covering consecutive video frames.

[0049] During training, a phased optimization strategy is adopted: first, the generator and discriminator are jointly trained using a degradation-clarity paired dataset, and the model's ability to perceive single degradation features is gradually improved through an alternating optimization strategy; then, an adaptive learning rate adjustment strategy is combined to ensure stable convergence and enhanced performance of the model under single degradation conditions such as low light, backlight, low resolution, and dynamic blur.

[0050] The loss of the generator is expressed by the following formula: ; in, Indicates the loss of the generator: This indicates degradation perception adversarial loss; This represents the multi-scale feature matching loss; This represents the loss due to physical constraints. This represents the frequency domain consistency loss; , , These represent the weight coefficients corresponding to the multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss, respectively. Their sensitivity and optimal range are determined through ablation experiments during training. The value ranges from 0.1 to 1.0. The value ranges from 0.01 to 0.5. The value ranges from 0.01 to 0.3.

[0051] The discriminator loss function is expressed by the following expression: ; in, Indicates discriminator loss; It is the discriminator's judgment result on the real sample; It is the discriminator's judgment result on the generated samples; It is the gradient penalty coefficient, with a value range of [1, 10]. It is a linear interpolation between real samples and generated samples.

[0052] By employing degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss to characterize the generator's loss, the model's ability to perceive single degradation features can be more effectively improved. Furthermore, combined with an adaptive learning rate adjustment strategy, the model can ensure stable convergence and enhanced performance under single degradation conditions such as low lighting, backlighting, low resolution, and dynamic blur, thereby obtaining a more accurate video processing model and achieving high-quality video frames after enhancement and optimization.

[0053] It should be noted that the above is only an optional example and should not be construed as a limitation of the present invention.

[0054] Next, in step S103, the video to be processed is input into the trained video processing model to obtain enhanced and optimized video data.

[0055] Specifically, when applying the model, the input is the surveillance video frame to be processed (which may contain degradation issues such as low light, backlight, low resolution, or motion blur), and the corresponding degradation type label (e.g., "low light"). The output is a high-quality video frame that has been enhanced and optimized, with visual quality (PSNR ≥ 30dB, SSIM ≥ 0.9) and detail reproduction close to the reference frame (HD 1920×1080 resolution). This allows for the acquisition of higher-quality playback video.

[0056] It should be noted that the above is only an optional example and should not be construed as a limitation of the present invention.

[0057] Next, in step S104, the obtained enhanced and optimized video data is spatiotemporally sliced ​​to form structured information storage, a bidirectional hash mapping table between the slice storage address and the time interval is established, and an index query process is executed through a retrieval mechanism to obtain the playback video corresponding to the user's request.

[0058] Specifically, based on video segmentation strategies such as H.264 / H.265 encoding, the video to be processed is spatiotemporally sliced ​​according to a preset time window (e.g., 5 seconds / slice). Distributed storage of video segments is implemented using, for example, an object storage system. At the same time, a metadata database is constructed to record structured information such as segment index, timestamp, and storage path.

[0059] A video segmentation strategy based on H.264 / H.265 encoding is adopted. The specific steps are as follows:

[0060] The video to be processed (including the enhanced and optimized video data obtained through the above steps) is spatiotemporally sliced, specifically by slicing the video into segments according to a preset time window (e.g., 5 seconds per segment). When the video frame rate is... At FPS, each video segment contains the following number of frames:

[0061] ; Each video segment is H.264 / H.265 encoded, and the encoded video segments are stored in the object storage system. The storage path for each video segment is represented as follows:

[0062] ; in, This represents the storage path of the i-th video segment; Indicates the underlying storage path; Indicates the first The start timestamp of each fragment; It is the file extension (e.g., .mp4).

[0063] The retrieval mechanism includes a pre-built caching mechanism. Based on this pre-built caching mechanism, a metadata database is constructed to record structured information such as the segment index, timestamp, and storage path of each video segment, forming the storage structure information of the pre-built caching mechanism. The metadata table is structured as follows:

[0064] ; in, This represents the structural information of the metadata table; It is the unique index of the i-th video segment. and These are the start and end timestamps of the i-th video segment, respectively. It is the storage path of the i-th video segment.

[0065] A fast retrieval algorithm based on time series is used to establish a bidirectional hash mapping table between fragmented storage addresses and time intervals. A pre-built caching mechanism accelerates the index query process, achieving millisecond-level video segment positioning capability.

[0066] Establish a bidirectional hash mapping table between the storage address and time interval of each video segment. The key of the bidirectional hash mapping table is the timestamp, and the value is the segment's storage path. The hash mapping table is represented as follows:

[0067] ; in, A bidirectional hash mapping table representing the storage address and time interval of each video segment; This represents the key corresponding to the timestamp of the i-th video segment; This represents the storage path of the i-th video segment.

[0068] The retrieval algorithm also includes a caching mechanism, which specifically accelerates the index query process through a pre-defined caching mechanism. The cache stores recently accessed shard indexes and storage paths, reducing disk I / O operations. The cache structure is represented as follows:

[0069] ; in, This indicates the cache storage of the recently accessed shard index and storage path, i.e., the cache; This represents the key corresponding to the timestamp of the i-th video segment; This represents the storage path of the i-th video segment. The cache can be stored by time, and the storage time of the cache can also be configured.

[0070] For the retrieval process, given a timestamp, the retrieval algorithm first searches in the cache. If the cache is not found, it then searches for the storage path of the corresponding video segment in a bidirectional hash table. The retrieval process is represented as follows:

[0071] ; This describes the process by which the retrieval algorithm searches for the corresponding video segment storage path based on a given timestamp t. The logic is to first search the cache; if a match is found, the cached path is returned; otherwise, the search continues in the bidirectional hash table. This indicates how to find the storage path of the video segment corresponding to timestamp t in the cache. If timestamp t exists in the cache, the corresponding path is returned directly, thus avoiding disk access and improving retrieval efficiency. This indicates how to find the storage path of the video segment corresponding to timestamp t in a bidirectional hash table. In the event of a cache miss, the hash table is used to quickly locate the actual storage location of the segment.

[0072] In a specific application example, it is applied to an interactive enhancement processing system, which includes a user-operable visual playback interface and an enhancement processing system that can interact with the visual playback interface.

[0073] Specifically, the visual playback interface allows users to select a time period and check the degradation type to be processed. The enhancement processing system automatically parses the request parameters (selected time period, checked degradation type) and triggers the CGAN inference engine under the corresponding conditions (specifically including low light, backlight, low resolution and image blur) to generate a video frame sequence with adaptive enhancement characteristics.

[0074] The visual playback interface includes the following components: a time selector, allowing users to select the start and end times; a degradation type selector, allowing users to check the degradation type to be processed (such as low light, backlight, low resolution, motion blur); and an enhancement button, used to trigger the enhancement processing flow.

[0075] For request parameter parsing and CGAN inference, when a user enters a request in the visual playback interface and clicks the enhancement button, the enhancement processing system automatically parses the request parameters and triggers the corresponding CGAN inference engine. Specifically, it parses the user-selected time period and degradation type to generate request parameters. ; in, It is a list of degradation types selected by the user.

[0076] For video segment retrieval, according to and Retrieve the corresponding video segment path: ; in, Indicates the storage path of the current video segment; Indicates the start time of the current video segment; Indicates the end time of the video segment; This describes the process by which the retrieval algorithm finds the corresponding video segment storage path based on a given timestamp t.

[0077] CGAN inference is performed on each video segment to generate an enhanced sequence of video frames. The CGAN inference process is represented as follows:

[0078] ; in, It is the first A sequence of frames from a video segment. It is the enhanced frame sequence.

[0079] The enhanced video frame sequence is re-encoded into a video file and displayed to the user on the interface. The display process is represented as follows:

[0080] ; in, This indicates that the enhanced video data or video file is displayed on the interface; It is the enhanced video data or video file.

[0081] A fast retrieval algorithm based on time series is established to create a bidirectional hash mapping table between the storage address of the segment and the time interval. The index query process is accelerated through a pre-built caching mechanism. If the timestamp t exists in the cache, the corresponding path is returned directly, thereby avoiding disk access and improving retrieval efficiency. In the case of cache miss, the actual storage location of the segment is quickly located through the hash table, which can achieve millisecond-level video segment location capability.

[0082] It should be noted that the above is only an optional example and should not be construed as a limitation of the present invention.

[0083] Compared with existing technologies, this invention constructs a video processing model based on conditional generative adversarial networks, trains the video processing model using the constructed training dataset to obtain a trained video processing model, and inputs the video to be processed into the trained video processing model to obtain enhanced and optimized video data, effectively improving the visual effect of the monitoring playback image.

[0084] Specifically, by introducing a conditional generative adversarial network (GAN), the generation strategy is dynamically adjusted to avoid over-enhancement in local areas while preserving and enhancing image details. This effectively improves the visual quality of surveillance playback images and the accuracy of subsequent analysis. A hybrid domain generator is used, combining spatial and frequency domain dual-path processing mechanisms. Through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules, local texture details are enhanced while noise is suppressed, ensuring that edges and details in surveillance playback images are clearly visible. Through a multi-granularity discriminator and an adaptive loss function system, more natural and higher-quality enhanced images of surveillance playback images are generated, reducing human intervention and improving the visual quality and realism of surveillance playback images.

[0085] Furthermore, the dynamic conditional coding module based on conditional generative adversarial networks can adapt to complex and ever-changing monitoring environments, effectively handle various degradation scenarios, and significantly improve the clarity and detail of monitoring playback videos.

[0086] Example 2 The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the embodiments of the apparatus of the present invention, please refer to the embodiments of the method of the present invention.

[0087] Figure 4 This is a schematic diagram illustrating an example of a monitoring playback processing device based on a conditional generative adversarial network according to the present invention. The following will refer to... Figure 4 The present invention describes a surveillance video playback device. This device is used to execute the surveillance video playback processing method described in the first aspect of the present invention.

[0088] like Figure 4 As shown, the monitoring playback processing device 400 includes an acquisition and processing module 410, a model building module 420, an output module 330, and a query and retrieval module 440.

[0089] In one specific implementation, the acquisition and processing module 410, based on a monitoring business scenario, acquires degraded video frame samples under multiple degradation conditions and reference frame samples corresponding to the business scenario, establishing a training dataset with degradation type and clarity pairing. The degraded video frame samples include video frame samples with the following degradation types: low-light type, backlight type, low-resolution type, and dynamic blur type. The model building module 420 constructs a video processing model based on a conditional generative adversarial network, and trains the video processing model using the constructed training dataset to obtain a trained video processing model. Specifically, during the training process, the degradation type is mapped to a conditional vector of a specified dimension and fused with the video to be processed as the model input to generate degradation-aware features; local texture details are enhanced and noise is suppressed through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules; hierarchical optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss; the output module 430 inputs the video to be processed into the trained video processing model to obtain enhanced and optimized video data. The query and retrieval module 440 performs spatiotemporal slicing of the obtained enhanced and optimized video data to form structured information storage, establishes a bidirectional hash mapping table between the slice storage address and the time interval, and executes the index query process through the retrieval mechanism to obtain the playback video corresponding to the user's request.

[0090] According to an optional implementation, the conditional generative adversarial network architecture maps degradation types to conditional vectors of a specified dimension, specifically 100 to 300, through a dynamic conditional coding module. By inputting the specified dimension condition vector, the degraded video frame sample, and the corresponding reference frame sample into the video to be processed, the following degradation perception features are obtained: degradation type features, local texture detail features, frequency domain features, multi-scale features, and physical constraint features.

[0091] According to an optional implementation, the spatial domain path of the dual-path processing mechanism employs variable convolution to calculate the offset of each sampling point in the deformable convolution relative to the regular grid position.

[0092] Specifically, it enhances low-level texture features, local structural features, and high-frequency details associated with the four degradation types.

[0093] According to an optional implementation, the loss of the generator is represented by the following expression: ; in, Indicates the loss of the generator: This indicates degradation perception adversarial loss; This represents the multi-scale feature matching loss; This represents the loss due to physical constraints. This represents the frequency domain consistency loss; , , These represent the weight coefficients corresponding to the multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss, respectively. The optimal range is determined by ablation experiments during training.

[0094] The value ranges from 0.1 to 1.0. The value can be 0.01, 0.5. The value ranges from 0.01 to 0.3.

[0095] According to the optional implementation method, the training dataset is established using the following expression: ; in, The training dataset representing the degradation-reference frame pairing; For the i-th degradation type of degraded video frame sample; Let H be the high-definition (1920×1080) reference frame sample corresponding to the degraded video frame sample of the i-th degradation type, where H represents the vertical number of pixels in the video frame and W represents the horizontal number of pixels in the video frame. and of , ; This is the label for the i-th degradation type.

[0096] According to the optional implementation method, in the traffic monitoring scenario, the segment duration is 5 seconds to 30 seconds.

[0097] In security monitoring scenarios, the segmentation time for key areas in specific residential communities and banks is 10 to 30 seconds; the segmentation time for key areas in shopping malls and public places is 1 to 2 minutes.

[0098] In specific environmental monitoring scenarios, segments of 5 to 10 minutes are used.

[0099] According to an optional implementation, a unique index identifier, i.e., a segment timestamp index, is generated for each video segment in the video segment sequence, and a bidirectional mapping table between the index identifier and each video segment is established. The bidirectional mapping table is used to perform video segment retrieval to obtain the playback video corresponding to the user's request.

[0100] It should be noted that, due to Figure 4 The monitoring playback processing method executed by the monitoring playback processing device based on conditional generative adversarial networks and Figure 1 The processing methods for the surveillance playback images in the examples are largely the same, therefore, the descriptions of the same parts have been omitted.

[0101] Compared with existing technologies, this invention constructs a video processing model based on conditional generative adversarial networks, trains the video processing model using the constructed training dataset to obtain a trained video processing model, and inputs the video to be processed into the trained video processing model to obtain enhanced and optimized video data, effectively improving the visual effect of the monitoring playback image.

[0102] Specifically, by introducing a conditional generative adversarial network (GAN), the generation strategy is dynamically adjusted to avoid over-enhancement in local areas while preserving and enhancing image details. This effectively improves the visual quality of surveillance playback images and the accuracy of subsequent analysis. A hybrid domain generator is used, combining spatial and frequency domain dual-path processing mechanisms. Through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules, local texture details are enhanced while noise is suppressed, ensuring that edges and details in surveillance playback images are clearly visible. Through a multi-granularity discriminator and an adaptive loss function system, more natural and higher-quality enhanced images of surveillance playback images are generated, reducing human intervention and improving the visual quality and realism of surveillance playback images.

[0103] Furthermore, the dynamic conditional coding module based on conditional generative adversarial networks can adapt to complex and ever-changing monitoring environments, effectively handle various degradation scenarios, and significantly improve the clarity and detail of monitoring playback videos.

[0104] Figure 5This is a schematic diagram of an embodiment of an electronic device according to the present invention.

[0105] like Figure 5 As shown, the electronic device is embodied in the form of a general-purpose computing device. There can be one or more processors working collaboratively. This invention also does not preclude distributed processing, meaning that processors can be distributed across different physical devices. The electronic device of this invention is not limited to a single entity, but can also be the sum of multiple physical devices.

[0106] The memory stores a computer-executable program, typically machine-readable code. The computer-readable program can be executed by the processor to enable the electronic device to perform the method of the present invention, or at least some steps of the method.

[0107] The memory includes volatile memory, such as random access memory (RAM) and / or cache memory, and may also be non-volatile memory, such as read-only memory (ROM).

[0108] Optionally, in this embodiment, the electronic device further includes an I / O interface for exchanging data with external devices. The I / O interface can represent one or more of several bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0109] It should be understood that Figure 5 The electronic device shown is merely one example of the present invention, and the electronic device of the present invention may also include elements or components not shown in the above examples. For example, some electronic devices also include display units such as displays, and some electronic devices also include human-computer interaction elements such as buttons and keyboards. Any electronic device capable of executing a computer-readable program in memory to implement the method of the present invention or at least some steps of the method can be considered as an electronic device covered by the present invention.

[0110] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software, or by combining software with necessary hardware. Therefore, as... Figure 6 As shown, the technical solution according to the embodiments of the present invention can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) or on a network, and includes several commands to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the above-described method according to the embodiments of the present invention.

[0111] The software product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0112] The computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with a command execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0113] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0114] The aforementioned computer-readable medium carries one or more programs, which, when executed by a device, enable the computer-readable medium to implement the data interaction method of this disclosure.

[0115] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0116] Through the description of the above embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions of the embodiments of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, portable hard drive, etc.) or on a network, including several commands to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of the present invention.

[0117] It should be noted that the above detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0118] In the detailed description above, reference has been made to the accompanying drawings, which form part of this document. In the drawings, similar symbols typically identify similar parts unless the context otherwise indicates otherwise. The illustrated embodiments described in the detailed specification, drawings, and claims are not intended to be limiting. Other embodiments may be used and other changes may be made without departing from the spirit or scope of the subject matter presented herein.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for processing surveillance playback images based on conditional generative adversarial networks, characterized in that, The method includes: Based on the monitoring business scenario, degraded video frame samples under multiple degradation conditions and reference frame samples of corresponding business scenarios are collected to establish a training dataset with degradation type and clarity pairing. The degraded video frame samples include video frame samples of the following degradation types: low light type, backlight type, low resolution type and dynamic blur type. A video processing model is constructed based on a conditional generative adversarial network (GAN). The model is trained using a constructed training dataset to obtain a trained video processing model. Specifically, during training, degradation types are mapped to conditional vectors of a specified dimension and fused with the video to be processed as model input to generate degradation-aware features. Local texture details are enhanced and noise is suppressed through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules. Hierarchical optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss, including: The loss of the generator is expressed by the following formula: in, Indicates the loss of the generator: This indicates degradation perception adversarial loss; This represents the multi-scale feature matching loss; Indicates physical constraint loss; This represents the frequency domain consistency loss; , , represents the weight coefficients corresponding to the multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss, respectively. During training, ablation experiments are used to determine the optimal range. The value ranges from 0.1 to 1.

0. The value can be 0.01 or 0.

5. The value ranges from 0.01 to 0.3; The video to be processed is input into the trained video processing model to obtain enhanced and optimized video data; The enhanced and optimized video data is spatiotemporally sliced ​​to form structured information storage. A bidirectional hash mapping table between the slice storage address and the time interval is established. An index query process is executed through a retrieval mechanism to obtain the playback video corresponding to the user's request.

2. The monitoring playback image processing method according to claim 1, characterized in that, During training, the degradation type is mapped to a conditional vector of a specified dimension and fused with the video to be processed as model input to generate degradation-aware features, including: The conditional generative adversarial network architecture maps degradation types to conditional vectors of specified dimensions, specifically 100 to 300 dimensions, through a dynamic conditional coding module. By inputting the specified dimension condition vector, the degraded video frame sample, and its corresponding reference frame sample into the video to be processed, the following degradation-perceived features are obtained: Degradation type features, local texture detail features, frequency domain features, multi-scale features, and physical constraint features.

3. The monitoring playback image processing method according to claim 2, characterized in that, The method enhances local texture details and suppresses noise through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules, including: Based on the spatial domain path of the dual-path processing mechanism, variable convolution is used to calculate the offset of each sampling point in the deformable convolution relative to the position of the regular grid. Specifically, it enhances low-level texture features, local structural features, and high-frequency details associated with the four degradation types.

4. The monitoring playback image processing method according to claim 1, characterized in that, include: The training dataset is created using the following expression: in, The training dataset representing the degradation-reference frame pairing; For the i-th degradation type of degraded video frame sample; Let H be the high-definition reference frame sample corresponding to the degraded video frame sample of the i-th degradation type, where H represents the vertical number of pixels in the video frame and W represents the horizontal number of pixels in the video frame. and of , ; This is the label for the i-th degradation type.

5. The monitoring playback image processing method according to claim 1, characterized in that, include: In traffic monitoring scenarios, the segment duration ranges from 5 seconds to 30 seconds; In security monitoring scenarios, the segmentation time for key areas in specific residential communities and banks is 10 to 30 seconds; the segmentation time for key areas in shopping malls and public places is 1 to 2 minutes. In specific environmental monitoring scenarios, segments of 5 to 10 minutes are used.

6. The monitoring playback image processing method according to claim 1, characterized in that, Further includes: Generate a unique index identifier, i.e., a segment timestamp index, for each video segment in the video segment sequence, and establish a bidirectional mapping table between the index identifier and each video segment; A bidirectional mapping table is used to perform video segment retrieval to obtain the playback video corresponding to the user's request.

7. A monitoring playback image processing device based on conditional generative adversarial networks, characterized in that, The monitoring playback image processing device, which performs the monitoring playback image processing method according to any one of claims 1 to 6, comprises: The acquisition and processing module, based on the monitoring business scenario, acquires degraded video frame samples under multiple degradation conditions and reference frame samples of the corresponding business scenario, and establishes a training dataset with degradation type and clarity pairing. The degraded video frame samples include video frame samples of the following degradation types: low light type, backlight type, low resolution type and dynamic blur type. The model building module, based on a conditional generative adversarial network, constructs a video processing model and trains it using the constructed training dataset to obtain a trained video processing model. Specifically, during training, degradation types are mapped to conditional vectors of a specified dimension and fused with the video to be processed as model input to generate degradation-aware features; local texture details are enhanced and noise is suppressed through deformable convolution, channel attention, fast Fourier transform, and adaptive filtering modules; and hierarchical optimization is achieved based on degradation-aware adversarial loss, multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss. The output module takes the video to be processed as input to the trained video processing model to obtain enhanced and optimized video data. The query and retrieval module spatiotemporally slices the enhanced and optimized video data to form structured information storage, establishes a bidirectional hash mapping table between the slice storage address and the time interval, and executes the index query process through the retrieval mechanism to obtain the playback video corresponding to the user's request.

8. The monitoring playback image processing device according to claim 7, characterized in that, include: The conditional generative adversarial network architecture maps degradation types to conditional vectors of specified dimensions, specifically 100 to 300 dimensions, through a dynamic conditional coding module. By inputting the specified dimension condition vector, the degraded video frame sample, and its corresponding reference frame sample into the video to be processed, the following degradation-perceived features are obtained: Degradation type features, local texture detail features, frequency domain features, multi-scale features, and physical constraint features.

9. The monitoring playback image processing device according to claim 7, characterized in that, include: The loss of the generator is expressed by the following formula: in, Indicates the loss of the generator: This indicates degradation perception adversarial loss; This represents the multi-scale feature matching loss; Indicates physical constraint loss; This represents the frequency domain consistency loss; , , represents the weight coefficients corresponding to the multi-scale feature matching loss, physical constraint loss, and frequency domain consistency loss, respectively. During training, ablation experiments are used to determine the optimal range. The value ranges from 0.1 to 1.

0. The value can be 0.01 or 0.

5. The value ranges from 0.01 to 0.3.

Images