Intelligent camera monitoring data storage and data enhancement processing method

Abstract

The invention relates to the technical field of monitoring, in particular to an intelligent camera monitoring data storage and data enhancement processing method, which effectively solves the image quality problem under a complex illumination condition through the combination of illumination invariant feature extraction and a weighted Gaussian probability model, especially the influence of dynamic shadow, uneven illumination and the like. According to the dynamic shielding synthesis mechanism, the dynamic shielding in a real scene is simulated through a Poisson fusion method based on a semantic segmentation result and a physically reasonable shielding object generation algorithm, the limitation of a conventional fixed template shielding method is avoided, the diversity and spatial rationality of shielding objects can be better reflected, and the real-time performance of the real scene is improved. The robustness of the model to shielding is improved; according to the invention, through combination of the cross-scale residual aggregation module and the gating channel-space attention mechanism, multi-scale feature reservation and deep fusion are realized, and detail information under small targets and complex illumination can be effectively captured in a complex monitoring scene.

Description

Technical Field

[0001] This invention relates to the field of surveillance technology, specifically to a method for storing and enhancing data in surveillance footage from intelligent cameras. Background Technology

[0002] With the widespread application of video surveillance technology, especially in public safety, traffic monitoring, and smart homes, the number of cameras and the amount of data generated are constantly increasing. This surveillance data faces significant challenges in terms of storage, processing, and analysis efficiency when providing functions such as real-time monitoring, video analytics, and event retrieval. Traditional surveillance data storage methods typically rely on centralized servers, which struggle to handle the storage demands of massive amounts of data and cannot guarantee efficient processing of real-time data streams. Meanwhile, existing image processing technologies perform poorly in complex environments such as changing lighting and dynamic occlusion, resulting in low accuracy in tasks like target detection and semantic segmentation in video surveillance, making it difficult to meet increasingly stringent monitoring requirements.

[0003] Furthermore, existing technologies have significant shortcomings in target tracking and image enhancement in dynamic scenes, particularly in real-time processing capabilities and robustness in complex environments. Traditional methods often rely on fixed algorithm templates, failing to flexibly adapt to changes in lighting and occlusion conditions in different scenarios. Due to the lack of effective multi-scale feature fusion and dynamic adaptation mechanisms, existing technologies are relatively weak in handling small targets, complex lighting, and occlusion situations, resulting in the system's monitoring accuracy and response speed failing to meet users' high standards in practical applications. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for intelligent camera monitoring data storage and data enhancement processing. By combining illumination-invariant feature extraction with a weighted Gaussian probability model, it effectively addresses image quality issues under complex lighting conditions, especially the effects of dynamic shadows and uneven illumination.

[0005] To achieve the above objectives, the following technical solution is adopted:

[0006] The method for data storage and data augmentation processing of intelligent camera surveillance includes the following steps:

[0007] S1. Capture diverse surveillance video streams using smart cameras, segment and label the captured raw videos, and obtain a training dataset;

[0008] S2. Store the acquired monitoring data;

[0009] S3. Enhance the preprocessing of monitoring data based on illumination perception segmentation, extract illumination-invariant features and generate semantic segmentation mask;

[0010] S4. Establish a dynamic occlusion synthesis mechanism, dynamically generate occlusion effects based on the segmentation mask, and embed real occlusion objects into the background area through Poisson fusion to simulate temporary occlusion in real scenes.

[0011] S5. Construct a data augmentation model, and process the data using the multi-scale feature-preserving encoder, attention-guided feature fusion, and edge-aware decoder reconstruction modules of the data augmentation model to generate high-quality reconstructed images.

[0012] S6. Output and store the enhanced image for subsequent advanced vision tasks.

[0013] Furthermore, in S1, video data is continuously recorded under different time periods and weather conditions when acquiring the training dataset to ensure that the data covers complex lighting and occlusion scenarios.

[0014] Furthermore, the original surveillance images and labeled masks in S2 are stored in a distributed file system or cloud storage platform, and an index database is established to associate the image and mask files to ensure fast retrieval and access.

[0015] Furthermore, S3 specifically refers to:

[0016] S31. Based on the original monitoring image, the dynamic range is compressed by logarithmic transformation, and Gaussian filtering is used to smooth the abrupt changes in illumination, thereby extracting illumination-invariant features.

[0017] S32. Based on illumination-invariant features, a weighted Gaussian probability model is used to calculate the semantic category of each pixel. By combining the prior weights of the categories and the difference measure between the features and the category center, a semantic segmentation mask is generated.

[0018] Furthermore, S4 specifically refers to:

[0019] S41. Based on the background semantic segmentation mask and the real occlusion image, the pose of the occlusion is adjusted by random transformation, and the transformed occlusion is seamlessly embedded into the background area using the Poisson fusion algorithm to generate a dynamic occlusion effect that conforms to the laws of physics.

[0020] S42. Based on randomly generated occlusion masks and dynamic occlusion images, the occlusion areas are combined with the original images through weighted fusion to generate an occlusion enhancement image.

[0021] Furthermore, S5 specifically refers to:

[0022] S51. A cross-scale residual aggregation module is adopted, which combines parallel downsampling paths and residual paths with a coordinate attention mechanism to achieve multi-scale feature preservation.

[0023] S52. By calculating the channel attention-gated and spatial attention-gated weighted encoder features and deeply fusing them with the decoder features, combined with segmentation semantic information, the focus on important channels is enhanced.

[0024] S53. By introducing segmentation boundary priors and multi-directional gradient loss, the reconstruction network is jointly optimized to improve the quality of edge detail restoration.

[0025] S54. A patch-level discriminator structure is adopted, and adversarial training is carried out by randomly cropping local image patches to enhance the ability to preserve local texture features.

[0026] S55. Based on reconstructed image, real image, edge loss, predictive segmentation mask, semantic segmentation mask and adversarial loss, a multi-task optimization objective is constructed by weighted combination of reconstruction loss, edge loss, segmentation loss and adversarial loss to balance pixel-level reconstruction quality, edge detail preservation, semantic consistency and local texture realism, so as to achieve end-to-end semantically consistent image enhancement.

[0027] S56. Iteratively train the data augmentation model.

[0028] Furthermore, S51 specifically refers to:

[0029] S511. Based on the input feature map and illumination-invariant features, downsampling is performed through convolution operation, and feature enhancement is performed by combining the scaled illumination-invariant features to generate a low-resolution feature map, so as to expand the receptive field and retain the robustness under low illumination.

[0030] S512. Based on the segmentation prior of input feature map and coordinate attention weighting, nonlinear transformation is performed through dense connected blocks to generate high-resolution residual features, so as to preserve detailed information and enhance the local texture representation capability.

[0031] S513. Based on downsampling features and upsampling residual features, multi-scale fusion features are generated by concatenating channels and fusing coordinate attention-weighted input features to simultaneously capture low-resolution semantic information and high-resolution detail information, thereby improving feature robustness.

[0032] S514, The encoder's final layer output yields the encoder features. It integrates multi-scale downsampling features and residual features, and enhances spatial perception capabilities through a coordinate attention mechanism.

[0033] Furthermore, S52 specifically refers to:

[0034] S521. Based on encoder features and semantic segmentation mask, channel information is extracted through global average pooling and global max pooling, and channel attention weights are generated by combining fully connected layers and Sigmoid activation.

[0035] S522. Based on encoder features and illumination-invariant features, spatial information is extracted through average pooling and max pooling, and spatial attention weights are generated by combining convolution operations and Sigmoid activation.

[0036] S523. Based on the encoder features weighted by channel attention gating and spatial attention gating, a fused feature is generated by adding it to the decoder features processed by depth-separable convolution.

[0037] Furthermore, S53 specifically refers to:

[0038] S531. Based on fusion features, segmentation boundary features and multi-directional gradient features, an image reconstruction network is used to generate an edge-enhanced reconstructed image.

[0039] S532. Based on the reconstructed image and the real image, the gradient difference is calculated through a multi-directional gradient operator and weighted by an edge weight mask to generate an edge loss.

[0040] Furthermore, S54 specifically refers to:

[0041] S541. Based on the reconstructed image and semantic segmentation mask, local image patches are generated through random cropping to ensure that sampling is focused on the target region;

[0042] S542. Based on real images and semantic segmentation masks, local image patches are generated through random cropping operations to ensure that sampling is focused on the target region;

[0043] S543. Based on real and fake patches, a patch-level discriminator is used for discrimination, and the adversarial loss is calculated by combining the gradient penalty term.

[0044] Compared with the prior art, the advantages of this invention are as follows:

[0045] 1. This invention effectively addresses image quality issues under complex lighting conditions, especially the effects of dynamic shadows and uneven lighting, by combining illumination-invariant feature extraction with a weighted Gaussian probability model.

[0046] 2. The dynamic occlusion synthesis mechanism of the present invention is based on semantic segmentation results and physically reasonable occlusion generation algorithm. It simulates dynamic occlusion in real scene through Poisson fusion method, avoids the limitations of conventional fixed template occlusion method, can better reflect the diversity and spatial rationality of occlusion, and improves the robustness of model to occlusion.

[0047] 3. This invention combines a cross-scale residual aggregation module with a gated channel-spatial attention mechanism to achieve multi-scale feature preservation and deep fusion, which can effectively capture detailed information of small targets and under complex lighting conditions in complex monitoring scenarios. Attached Figure Description

[0048] Appendix Figure 1 This is a flowchart of the present invention;

[0049] Appendix Figure 2 This is a schematic diagram illustrating the impact of the dynamic occlusion synthesis mechanism on target detection performance in this invention.

[0050] Appendix Figure 3 A schematic diagram illustrating the feature selection effects of different attention mechanisms in complex scenes. Detailed Implementation

[0051] This invention proposes an intelligent camera monitoring data storage and data enhancement processing system, the main steps of which are as follows:

[0052] S1. Obtaining the training dataset

[0053] Deploy smart cameras in real-world surveillance scenarios to capture diverse surveillance video streams;

[0054] During the specific data collection process, cameras were installed in typical monitoring areas such as traffic intersections and public places to continuously record video data under different time periods (such as daytime, nighttime, and dusk) and different weather conditions (such as sunny days, rainy days, and foggy days) to ensure that the data covers complex lighting and occlusion scenarios.

[0055] The captured raw video is segmented into image frames and stored at high resolution (e.g., H×W×3).

[0056] The data annotation stage adopts a pixel-level semantic segmentation annotation method. Professional annotators use annotation tools (such as LabelMe or CVAT) to label each pixel in each image with a category and generate a corresponding semantic segmentation mask.

[0057] The annotation categories include vehicles, pedestrians, roads, and shadows. For example, in a traffic monitoring scenario, the number of semantic categories K=4, corresponding to the vehicle, pedestrian, road, and shadow categories, respectively.

[0058] The annotation results are stored in the form of a One-Hot encoded mask, where each pixel position is represented by a K-dimensional vector, where 0 indicates that it does not belong to the class and 1 indicates that it belongs to the class, thus providing accurate semantic priors for model training.

[0059] S2, Monitoring Data Storage

[0060] The monitoring data storage process is responsible for the efficient organization and management of the collected raw monitoring images and corresponding semantic segmentation masks;

[0061] During storage, the original surveillance images and labeled masks are stored in a distributed file system or cloud storage platform respectively, and an index database is established to associate the image and mask files to ensure fast retrieval and access.

[0062] The storage system employs a tiered architecture, placing frequently accessed training data in a high-speed storage tier (such as SSDs) and archiving low-frequency data in a low-cost storage tier (such as HDDs) to balance performance and cost.

[0063] Data is preprocessed before storage, including format standardization (such as converting to a uniform resolution) and metadata recording (such as timestamps, scene types, and lighting conditions), so that subsequent data augmentation modules can call them as needed.

[0064] The role of storage is to provide stable and scalable data support for model training and evaluation, while ensuring data security and integrity through redundant backups and fault tolerance mechanisms, and avoiding the impact of data loss or damage on system reliability.

[0065] In addition, the storage system supports dynamic data stream processing, allowing real-time monitoring of data inflows and synchronous storage, providing a continuous data source for online enhancements and model updates.

[0066] S3. Light-sensing-based segmentation enhancement preprocessing

[0067] In intelligent camera surveillance systems, complex surveillance images are often affected by uneven lighting, low contrast, and dynamic shadows, resulting in a significant decrease in image quality. Conventional processing methods use histogram equalization to improve contrast, but this amplifies image noise and destroys spatial continuity, making it difficult to effectively improve the accuracy of subsequent segmentation tasks.

[0068] This invention eliminates the influence of illumination and enhances segmentation boundaries through illumination-invariant feature extraction and semantic segmentation mask generation. The specific steps are as follows:

[0069] 1) Illumination-invariant feature extraction

[0070] Based on the original surveillance images, the dynamic range is compressed through logarithmic transformation, and Gaussian filtering is used to smooth abrupt changes in illumination, thereby extracting illumination-invariant features to eliminate interference from illumination changes and preserve the essential texture of objects, represented as:

[0071]

[0072] In the formula, Represents illumination-invariant features, with dimension . Logarithmic transformation is used to compress the dynamic range, Gaussian filtering is used to smooth abrupt changes in lighting, shadow interference is eliminated, and the essential texture of the object is preserved.

[0073] Represents the original surveillance image, with dimensions of These are the raw surveillance images directly captured by the camera; Indicates the image height; Indicates the image width; This represents a zero-preserving constant to avoid undefined values ​​in logarithmic operations; the preferred value is [value to be filled in]. ; This indicates logarithmic operations, with the default base being the natural constant. This term is used to compress the dynamic range of an image; This represents a Gaussian filter kernel used for smoothing images. The standard deviation of the Gaussian kernel is preferably set to [value missing]. ; This represents the convolution operation.

[0074] 2) Semantic segmentation mask generation

[0075] Based on illumination-invariant features, a weighted Gaussian probability model is used to calculate the semantic category of each pixel. By combining prior category weights and a difference metric between features and category centers, a semantic segmentation mask is generated to improve segmentation boundary accuracy and address the issue of large fluctuations in intra-class features under uneven illumination. This is represented as:

[0076]

[0077] In the formula, Represents a semantic segmentation mask with dimension . It is a One-Hot encoding mask. This indicates that the pixel does not belong to this class. This indicates that it belongs to this class;

[0078] This represents the number of semantic categories, including categories such as foreground, background, and shadow, as in traffic monitoring. This includes categories such as vehicles, pedestrians, roads, and shadows; This indicates that the category index that maximizes the internal expression is used to determine the category to which each pixel belongs; Represents a semantic category index; Indicates the first Prior weights for classes are used to balance the probabilities of different classes. Specifically, they are calculated using class frequencies based on training data statistics, expressed as follows: ; For the first The number of pixels in the class; This represents the total number of pixels. Represents the natural exponential function; This represents the L2 norm, also known as the Euclidean norm. Indicates the first The mean of class features; Indicates the first Standard deviation of class features.

[0079] In practical implementation, and The segmentation mask is estimated from the training data, specifically using the labeled segmentation mask for the th... For each category, calculate the illumination-invariant features of all pixels belonging to that category. The mean and standard deviation are obtained from the above.

[0080] It should be noted that, The term represents the square of the Mahalanobis distance, used to measure the difference between pixel features and class centers, taking into account the influence of feature distribution variance. The judgment is relaxed when the time is large, thereby solving the problem of large fluctuations in intra-class features under uneven lighting and improving the accuracy of distinguishing shadows from objects.

[0081] It should also be noted that the weighted Gaussian probability model uses Mahalanobis distance to measure the difference between pixel features and the class center. It not only considers the difference in feature values, but also takes into account the influence of the variance of the class feature distribution. When the feature variance of a certain class is large, such as due to changes in illumination causing large fluctuations in intra-class features, the Mahalanobis distance will automatically relax the judgment threshold, thereby avoiding misclassification of pixels that belong to that class.

[0082] In one embodiment, the semantic segmentation accuracy under different lighting conditions is compared to verify the robustness and adaptability of the proposed technology in various complex lighting environments, particularly its performance in challenging surveillance scenarios such as nighttime, rainy days, and foggy days. The experiment compares four different image processing methods, including conventional histogram equalization, the vision-based Retinex algorithm, conventional semantic segmentation networks, and the complete method proposed in this invention. The experimental configuration simulates six common lighting conditions in real-world surveillance scenarios, ranging from ideal sunny daytime to extremely challenging nighttime, rainy, and foggy conditions. The same number of test samples were used for evaluation under each condition. The experimental results clearly show that the proposed method maintains the highest semantic segmentation accuracy under all lighting conditions, especially in poor lighting conditions such as dusk, nighttime, rainy days, and foggy days, where its advantages are even more pronounced. Conventional histogram equalization methods, due to amplifying image noise and disrupting spatial continuity, exhibit the most significant performance degradation under complex lighting conditions. While the Retinex algorithm improves the problem of uneven lighting to some extent, it lacks effective utilization of semantic information, resulting in limited performance improvement. While conventional segmentation networks possess some learning capabilities, they are not specifically optimized for changes in illumination, resulting in significant shortcomings under low-light conditions. Our proposed technique, however, effectively eliminates the interference of illumination variations and preserves the essential texture features of objects through illumination-invariant feature extraction and semantic segmentation mask generation. Therefore, it maintains stable high performance under various illumination conditions.

[0083] S4, Dynamic Occlusion Synthesis Mechanism

[0084] In surveillance scenarios, targets are often partially obscured by temporary obstacles such as pedestrians or vehicles, resulting in missing target information. Conventional data augmentation methods use fixed occlusion templates for simulation, but this can cause a shift in the distribution of training data and fail to effectively reflect the dynamic diversity of occlusion in real-world scenarios.

[0085] This invention dynamically generates physically plausible occlusion effects based on semantic segmentation results, and naturally embeds real occluders into the image background region through Poisson fusion. The specific steps are as follows:

[0086] 1) Dynamic Occlusion Generation

[0087] Based on a background semantic segmentation mask and real occlusion images, the pose of the occluders is adjusted by random transformation, and the transformed occluders are seamlessly embedded into the background region using a Poisson fusion algorithm to generate a dynamic occlusion effect that conforms to physical laws, simulating temporary occlusion in real-world scenes, as shown below:

[0088]

[0089] In the formula, This indicates that a random transformation, such as rotation or scaling, is applied to the image of the actual occluded object. These are transformation parameters; Based on semantic segmentation mask The generated binary mask ensures that the occluder is placed only in the background area, with a dimension of . ; The occlusion image is a dynamically generated image used to simulate temporary occlusion in a real-world scene, with dimensions of [dimension value missing]. ; Represents the background semantic segmentation mask, with dimensions of It is a binary mask, where 1 represents the background region and 0 represents the non-background region. This is achieved by analyzing the semantic segmentation mask. The background category is obtained by extracting it. This represents the Poisson fusion operator, which achieves seamless fusion between images by solving the Poisson equation; This represents element-wise multiplication. Represents real-world occlusion images, sourced from a predefined database of occlusion images, including common occlusions such as tree branches and billboards, used to simulate temporary occlusion in real-world scenes.

[0090] In one implementation, the transformation parameter θ is used to control the transformation of the image of the real occluder. The random geometric and photometric transformations are used to simulate the diversity and randomness of occlusions in real-world scenes. θ is a parameter vector that includes the rotation angle. and scaling factor Among them, rotation angle Random sampling causes a change in the orientation of the obstruction, scaling factor Controlling the fluctuation of the size of the occluder by analyzing the actual image of the occluder. Apply rotation and scaling transformations to generate the transformed spatial coordinates.

[0091] In practical implementation, The project performs a random transformation on the real occluded image, obtaining the transformed occluded image as the source image, and uses the background semantic segmentation mask. The corresponding original image region is used as the target image. The Poisson fusion operator generates a fused image by minimizing the gradient difference between the source and target images within the fusion region, thereby preserving the texture details of the occluded objects and naturally adapting to background lighting and color changes. The fusion region is based on a background semantic segmentation mask. The mask identifies background and non-background regions in the image. During the Poisson fusion process, the background semantic segmentation mask... It is used as a binary mask to specify the fusion boundary between the source image and the target image.

[0092] 2) Occlusion Image Synthesis

[0093] Based on randomly generated occlusion masks and dynamic occlusion images, a weighted fusion method is used to combine the occluded regions with the original image to generate an occlusion enhancement image. This simulates the realistic occlusion effect while maintaining the natural appearance of the image, and is represented as follows:

[0094]

[0095] In the formula, This represents an occlusion enhancement image with dimension . It includes dynamically generated occlusions to simulate realistic occlusion effects while maintaining the natural appearance of the image; This represents a matrix of all ones with the same dimension as the randomly generated occlusion mask. This represents a randomly generated occlusion mask with dimensions of . , is used to control the position and shape of the occluded area, where 1 represents the occluded area and 0 represents the unoccluded area; This indicates element-wise multiplication.

[0096] It should be noted that combining semantic priors with random transformations, and using a background semantic segmentation mask... Strictly limit occlusions to blending only into the background area to avoid interfering with foreground targets. Meanwhile, The Poisson fusion operator, by solving the gradient minimization problem, adapts to changes in background illumination while preserving the texture details of occluded objects, generating visually coherent occlusion effects. Through dynamic generation and semantic guidance, it simulates the randomness and spatial plausibility of realistic occlusions, significantly improving the diversity of training data and the model's generalization ability. Furthermore... Based on segmentation mask generation, the placement of occlusion objects does not damage the structure of key targets, which is difficult to achieve with conventional random cropping or fixed templates. The introduction of semantic prior makes the occlusion synthesis more in line with real physical constraints, thus producing a more realistic enhancement effect in complex monitoring scenarios.

[0097] In one embodiment, the impact of the dynamic occlusion synthesis mechanism on target detection performance is analyzed to verify the effectiveness of the proposed dynamic occlusion synthesis mechanism in dealing with different degrees of occlusion. The experiment set six different occlusion rate levels, from no occlusion to severe occlusion, and compared four different data processing strategies, including a basic method without any data augmentation, a conventional method using random cropping augmentation, an improved method using a fixed occlusion template, and the complete solution of this technique. In the experimental configuration, the occlusion rate gradually increased from 0 to 50%, simulating the actual situation of targets being occluded to different degrees in real-world surveillance scenarios. The experimental results show that as the occlusion rate increases, the detection performance of all methods decreases, but the decrease in the proposed method is significantly smaller than that of the other comparative methods. Under high occlusion rate conditions, this technique can still maintain relatively high detection accuracy, fully demonstrating the effectiveness of the dynamic occlusion synthesis mechanism. Although the conventional random cropping method can improve the robustness of the model to occlusion to some extent, its effect is limited due to a lack of physical plausibility. Although the fixed occlusion template method is an improvement over random cropping, it cannot reflect the diversity and randomness of occlusion objects in real-world scenes. This technology, through a Poisson fusion mechanism based on semantic segmentation results, can generate dynamic occlusion effects that conform to physical laws, significantly improving the model's generalization ability in occlusion scenarios.

[0098] S5. Building Data Augmentation Models

[0099] S501 Multi-Scale Feature Preservation Encoder

[0100] In complex monitoring scenarios, the scale of targets varies greatly. Conventional convolutional encoders lose detailed features of small targets due to multiple downsampling operations, resulting in reduced feature map resolution and difficulty in effectively preserving the structural information of tiny targets.

[0101] This invention employs a cross-scale residual aggregation module, which combines parallel downsampling paths with residual paths and a coordinate attention mechanism to achieve multi-scale feature preservation. The specific steps are as follows:

[0102] 1) Downsampling feature extraction

[0103] Based on the input feature map and illumination-invariant features, downsampling is performed through convolution, and feature enhancement is performed by combining the scaled illumination-invariant features to generate a low-resolution feature map. This expands the receptive field and preserves robustness under low illumination, as shown below:

[0104]

[0105] In the formula, Indicates the encoder's first... The downsampled feature map of the layer has a dimension of , used to provide low-resolution features; This represents the activation function of the linear rectified unit with leakage, also known as the LeakyReLU activation function. The negative slope coefficient is preferably set to 0.2, which is used to retain a small amount of gradient flow in the negative region to avoid neuron death, especially to enhance nonlinear expression in low-light features. This indicates a batch normalization operation; This represents an image scaling operation, which adjusts the feature map size using bilinear interpolation. In the item, Indicates the characteristic of invariant illumination Size reduced to the original This is multiplied by a factor of 1 to match the resolution of the downsampled feature map; This represents element-wise addition. It is the encoder number The number of channels in the layer; Indicates the encoder's first... The downsampled feature map of the layer, when hour, This represents the input feature map of the encoder, i.e., illumination-invariant features. ; For the encoder's layer index; Indicates the kernel size as Convolution operations; This represents the convolution stride.

[0106] 2) Residual Feature Extraction

[0107] Based on the segmentation prior of the input feature map and coordinate attention weighting, a nonlinear transformation is performed through dense connected blocks to generate high-resolution residual features, which preserves detail information and enhances local texture representation, as shown below:

[0108]

[0109] In the formula, Indicates the encoder's first... The residual feature map of the layer retains high-resolution detail information through dense connected blocks and integrates coordinate attention-weighted segmentation priors, so that the feature map has rich local texture and boundary information in the spatial dimension, which helps to alleviate the problem of detail loss during downsampling. This indicates the coordinate attention module; Indicates the kernel size as Convolution operations; This represents a densely connected block, consisting of multiple convolutional layers, with the output of each layer connected to subsequent layers via channel splicing.

[0110] In its implementation, the coordinate attention module enhances the sensitivity of features to spatial location by decomposing channel attention into spatial encodings in both horizontal and vertical directions. Specifically, it first processes the input feature map... Global average pooling is performed in both the height and width directions. Attention weights are then generated by sharing a fully connected layer and a sigmoid activation. Finally, these weights are multiplied with the original features to output a spatially enhanced feature map.

[0111] In its implementation, a densely connected block consists of multiple convolutional layers. The input of each layer includes the concatenation of the outputs of all preceding layers. Specifically, its input is defined... for Define the dense connection block. The output of the layer is , represented as , It serves as a layer index for densely connected blocks, promoting feature reuse and gradient flow through dense connections, and enhancing nonlinear transformation capabilities.

[0112] 3) Cross-scale feature aggregation

[0113] Based on downsampling features and upsampling residual features, multi-scale fusion features are generated by concatenating and fusing coordinate attention-weighted input features through channel concatenation. This simultaneously captures low-resolution semantic information and high-resolution detail information, improving feature robustness. This is represented as:

[0114]

[0115] In the formula, Indicates the encoder's first... The output feature map of the layer has a dimension of It represents the features after multi-scale fusion, has rich semantic and detailed information, and preserves the structure of small targets through cross-scale aggregation; The number of channels for the residual feature; Indicates a channel splicing operation; This indicates a bilinear upsampling operation used to adjust the feature map size; Indicates the upsampling ratio; Indicates the encoder's first... The layer's coordinate attention module is used to enhance the spatial awareness of features.

[0116] In practical implementation, The first step is to analyze the feature map. Global average pooling is performed along both the height and width directions to obtain feature maps in two directions. Then, a spatial attention map is generated through convolution and activation functions, and finally... Element-wise multiplication outputs features with enhanced spatial awareness.

[0117] It should be noted that the cross-scale residual aggregation module is used to integrate the encoder's first... Downsampled feature map of the layer Features of upsampling residuals The items are concatenated into channels and then fused with coordinate attention-weighted input features. The term, through parallel paths, simultaneously captures low-resolution semantic information and high-resolution detail information, enabling the encoder to... Output feature map of the layer While preserving the structure of small targets, it also possesses rich contextual semantics. In particular, the dense connective blocks in the residual path enhance the representation of local textures through multi-layer feature reuse. The method enhances the perception of target boundaries through spatial location encoding, which can effectively improve feature robustness under complex lighting conditions.

[0118] 4) The encoder features are obtained from the output of the last layer of the encoder. By integrating multi-scale downsampling features and residual features, and enhancing spatial perception through a coordinate attention mechanism, it can both preserve low-resolution semantic information and integrate high-resolution details, thus effectively coping with scale changes and illumination interference in complex monitoring scenarios and avoiding the loss of small target features.

[0119] S502, Attention-Guided Feature Fusion

[0120] Conventional skip connections ignore the semantic differences between encoder and decoder features, leading to noise during feature fusion and reducing the quality of reconstructed images, especially in complex monitoring scenarios such as rain and fog where the false activation rate is high.

[0121] This invention introduces a gated channel-spatial attention mechanism. By calculating channel attention-gated and spatial attention-gated weighted encoder features and deeply fusing them with decoder features, combined with segmentation semantic information, it enhances the focus on important channels. The specific steps are as follows:

[0122] 1) Channel attention gating calculation

[0123] Based on encoder features and semantic segmentation masks, channel information is extracted through global average pooling and global max pooling. This is combined with fully connected layers and sigmoid activation to generate channel attention weights, thereby enhancing the feature responses of important channels and incorporating semantic priors. This is represented as follows:

[0124]

[0125] In the formula, Represents the channel attention gating weights, dimensions, and encoder features. The number of channels is the same, used to weight the feature importance of different channels; This indicates a weighted operation along the channel dimension. Item representation for semantic segmentation mask Weighting is performed along the channel dimension, through Convolution generates attention priors; This represents the Sigmoid activation function, which maps the input to... interval; This represents a global average pooling operation, which averages the spatial dimensions of the feature map and outputs a vector along the channel dimension. This represents a global max pooling operation, which takes the maximum value of the spatial dimension of the feature map and outputs a vector of the channel dimension. This represents the weight matrix of the first fully connected layer, which are trainable parameters; This represents the weight matrix of the second fully connected layer, which are trainable parameters; This indicates a concatenation operation, also known as a channel concatenation operation, used to merge two pooling features.

[0126] In practical implementation, The item implements the weighted operation of the channel dimension through a Convolutional layers process semantic segmentation masks ,Will The number of channels is adjusted to match the encoder features. Consistent, then the Sigmoid activation function is applied to generate channel attention weights, which are consistent with... Element-wise multiplication is performed to embed semantic prior information.

[0127] 2) Spatial attention gating computation

[0128] Based on encoder features and illumination-invariant features, spatial information is extracted through average pooling and max pooling, and spatial attention weights are generated by combining convolution operations and sigmoid activation to enhance feature responses at important spatial locations and improve boundary sensitivity, as expressed below:

[0129]

[0130] In the formula, Representing spatial attention gating weights, dimensions and encoder features They have the same spatial dimension and are used to weight the importance of features at different spatial locations; This represents a spatial transformation operation. The term characterizes the illumination-invariant features. Perform spatial transformations and extract edge enhancement features through convolutional layers; Indicates the kernel size as Convolution operations; This represents the average pooling result of the encoder features, achieved by applying the encoder features... The average of all channels at each location is obtained; This represents the max pooling result of the encoder features, achieved by applying the encoder features... The maximum value is obtained by taking the values ​​of all channels at each position.

[0131] In practical implementation, The item implements spatial transformation operations, through a Illumination-invariant features of convolutional layers Edge enhancement is performed to extract its texture features, and then bilinear interpolation is used to adjust the size to match the spatial dimension of the encoder features. Finally, it is fused with the pooling result to improve the boundary sensitivity of spatial attention.

[0132] 3) Gated feature fusion

[0133] The encoder features, weighted by channel attention gating and spatial attention gating, are added to the decoder features processed by depthwise separable convolution to generate fused features. This enhances key information in both channel and spatial dimensions and achieves deep fusion, as shown below:

[0134]

[0135] In the formula, The fused feature map represents the deep fusion of encoder and decoder features after gated weighting, while enhancing key information in the channel and spatial dimensions. This indicates a depthwise separable convolution operation, used to reduce computation and extract decoder features; Decoder features, dimensions, and encoder features same.

[0136] In practical implementation, The term implements depthwise separable convolution operations, first by processing the decoder features. Perform channel-by-channel Convolution, and then through Convolution adjusts the number of channels to reduce computational cost while preserving spatial information.

[0137] In practical implementation, decoder features Derived from the output of the previous layer of the decoder network, the decoder consists of multiple upsampling blocks, each containing upsampling and convolution operations. This is the current feature map of the decoder before attention fusion; specifically, the decoder uses the encoder's output features. As the initial input, spatial resolution is recovered through layer-by-layer upsampling.

[0138] It should be noted that a gated channel-spatial attention mechanism is adopted, which uses channel attention to gate weights. Spatial attention gating weights Weighted encoder features Channel attention is integrated into the semantic segmentation mask. Semantic priors, spatial attention combined with illumination-invariant features By dynamically adjusting the feature importance through multi-source information fusion, the texture information can not only enhance the feature response of important channels and spatial locations, but also suppress background noise and interference from irrelevant areas.

[0139] S503, Edge-Aware Decoder Reconstruction

[0140] Conventional decoders are prone to producing blurred boundaries during image reconstruction, failing to accurately recover the detailed structures in surveillance images, resulting in a decline in the quality of reconstructed images, especially in the target edge region.

[0141] This invention improves the quality of edge detail restoration by introducing segmentation boundary priors and multi-directional gradient loss to jointly optimize the reconstruction network. The specific steps are as follows:

[0142] 1) Edge-enhanced image reconstruction

[0143] Based on fusion features, segmentation boundary features, and multi-directional gradient features, an image reconstruction network is used to generate an edge-enhanced reconstructed image to preserve clear boundaries and texture details, as shown below:

[0144]

[0145] In the formula, Represents the reconstructed image, with dimensions of By enhancing edge priors, the clear boundaries and texture details of the target are preserved; This represents a deconvolutional reconstruction network, which consists of multiple deconvolutional layers. Indicates characteristics that are invariant to illumination Multi-directional gradient extraction is performed by calculating and concatenating the gradients in the horizontal, vertical, and diagonal directions using the Sobel operator. This indicates a channel concatenation operation, used to connect different features along the channel dimension; The segmentation boundary features are extracted from the segmentation mask using the Canny edge detection algorithm, with a dimension of [missing information]. It is used to provide edge prior information.

[0146] In its implementation, the deconvolutional reconstruction network consists of multiple deconvolutional layers. Each layer includes deconvolution, batch normalization, and LeakyReLU activation, progressively upsampling the feature maps to the original image size, and finally passing them through a... The convolutional layer outputs the reconstructed image.

[0147] In practical implementation, segmentation boundary features Using the Canny edge detection algorithm to segment the mask Extract from, specifically, first of all Convert to a binary image, then apply the Canny operator to detect edges, generating a single-channel binary mask with dimensions [not specified]. The value is 1 for edge positions and 0 for non-edge positions.

[0148] 2) Multi-directional edge loss calculation

[0149] Based on the reconstructed image and the real image, gradient differences are calculated using a multi-directional gradient operator and weighted by an edge weight mask to generate an edge loss. This strengthens the constraints on edge regions and improves the accuracy of detail recovery, as expressed below:

[0150]

[0151] In the formula, This represents edge loss, characterizing the difference between the reconstructed image and the real image in multi-directional gradients. It strengthens the constraints on edge regions, avoids blurring of reconstructed image boundaries, and improves visual quality. This represents the L1 norm, used to calculate the absolute value of the gradient difference; This indicates that the segmentation boundary features are binarized to generate an edge weight mask; Indicates the first Sobel gradient operators in each direction are used to calculate the gradient of the image in that direction; The real images are derived from high-quality surveillance images in the training dataset and serve as a standard reference for reconstructing images. They are used to calculate the loss function to optimize the network. This represents element-wise multiplication.

[0152] Indicates the gradient direction index; Indicates the first Edge region masks in each direction are obtained by thresholding the gradient magnitude in the corresponding direction, and are used to calculate the loss for focusing on the edge region.

[0153] In specific implementation, the first The Sobel gradient operator in each direction applies a gradient to the reconstructed image. Processing is performed on the input reconstructed image. Convolution is performed using a Sobel kernel, with gradient direction indexing. The Sobel kernels correspond to the horizontal, vertical, 45° diagonal, and 135° diagonal directions, respectively.

[0154] In specific implementation, the first Edge region mask in each direction The gradient magnitude in the corresponding direction is obtained by thresholding, and is represented as follows: ,in For indicator functions, This is a preset threshold, such as 0.1, used for loss calculation focusing on edge regions.

[0155] It should be noted that segmentation boundary features are used. As a segmentation boundary prior, multi-source edge features are fused, and gradient differences are calculated in four directions using a loss function. This edge prior guides the reconstruction network to focus on key boundary regions, while multi-directional gradient loss strengthens the constraints on edge structures, ensuring high consistency in target contours and texture details in the reconstructed image. Furthermore... Item fusion illumination invariant features The gradient information can enhance the robustness of the reconstruction process to changes in illumination and effectively restore clear edges under dynamic shadow interference.

[0156] S504, Adverse Local Enhancement

[0157] Global adversarial training tends to weaken local texture details in monitoring image enhancement, resulting in poor performance of reconstructed images in fine-grained features, especially in low-light conditions such as nighttime, where the texture similarity index is low.

[0158] This invention employs a patch-level discriminator structure, which generates local image patches through random cropping for adversarial training, thereby enhancing the ability to preserve local texture features. The specific steps are as follows:

[0159] 1) Local patch sampling from reconstructed images

[0160] Based on the reconstructed image and semantic segmentation mask, local image patches are generated through random cropping to ensure that sampling is focused on the target region, thereby providing false patch input and enhancing the realism of local textures, as shown below:

[0161]

[0162] In the formula, This represents a fake patch sampled from the reconstructed image, with a dimension of [missing information]. , representing the local texture region of the reconstructed image;

[0163] This indicates a random cropping operation, which randomly selects a crop from the input image. The area of ​​pixels;

[0164] 2) Local patch sampling from real images

[0165] Based on real images and semantic segmentation masks, local image patches are generated through random cropping to ensure that sampling focuses on the target region, providing real patch inputs and serving as positive examples for the discriminator, as shown below:

[0166]

[0167] In the formula, Represents a true patch sampled from a real image, with dimension . , representing local texture regions of a real image.

[0168] 3) Calculation of patch resistance loss

[0169] Based on real and fake patches, a patch-level discriminator is used for discrimination, and an adversarial loss is calculated by combining a gradient penalty term to improve the local texture realism of the reconstructed image and stabilize the training process, as expressed in:

[0170]

[0171] In the formula, The adversarial loss represents the degree of competition between the generator and the discriminator on local patches. By minimizing the generator loss and maximizing the discriminator loss, the realism of local texture in the reconstructed image is improved.

[0172] This indicates the expected operation, which calculates the average value on a batch of data; This indicates that the discriminator identifies true patches sampled from real images. The output gradient is used for the gradient penalty term, and the gradient norm is constrained to stabilize adversarial training. This is the gradient penalty coefficient, preferably 10, used to stabilize adversarial training and avoid mode collapse; This represents a patch-level discriminator, which uses a multi-layer convolutional structure to determine whether an input patch is real or fake.

[0173] In one implementation, the patch-level discriminator structure consists of multiple convolutional layers, each containing convolution, batch normalization, and LeakyReLU activation, and finally outputs probabilities through a fully connected layer and a sigmoid function.

[0174] It should be noted that, The term represents the output probability of the discriminator for true patches; the closer it is to 1, the better the discrimination effect. The term represents the output probability of the discriminator for false patches; the closer it is to 0, the better the discrimination effect. The term represents the loss term of the discriminator for true patches; maximizing this value improves the discriminative ability. The term represents the loss term of the discriminator for false patches; maximizing this value improves the discrimination ability.

[0175] It should be noted that the patch-level discriminator is trained stably by incorporating a gradient penalty term. This forces the generator to match the true distribution on fine-grained textures through adversarial learning of local patches. This approach ensures that adversarial training focuses on the foreground target and avoids interference from background noise, thereby improving the model's generalization ability in complex scenarios.

[0176] S505, Multi-task Joint Optimization

[0177] A single reconstruction loss function ignores semantic consistency in complex surveillance image enhancement, resulting in reconstructed images with high pixel-level similarity but incomplete semantic information, which affects the performance of subsequent advanced vision tasks.

[0178] This invention, based on reconstructed image, ground truth image, edge loss, predictive segmentation mask, semantic segmentation mask, and adversarial loss, constructs a multi-task optimization objective by weightedly combining reconstruction loss, edge loss, segmentation loss, and adversarial loss. This objective balances pixel-level reconstruction quality, edge detail preservation, semantic consistency, and local texture realism, achieving end-to-end semantically consistent image enhancement, as expressed below:

[0179]

[0180] In the formula, The total loss represents the objective function of multi-task joint optimization, balancing pixel-level reconstruction, edge preservation, semantic consistency, and local realism, and guiding the network to learn end-to-end through weighted combination.

[0181] The weighting coefficients representing the reconstruction loss control the importance of pixel-level reconstruction, and the preferred values ​​are [values ​​to be filled in]. ;

[0182] The weighting coefficients representing edge loss control the importance of preserving edge details; the preferred values ​​are [values ​​to be filled in]. ;

[0183] The weights representing the segmentation loss control the importance of semantic consistency, and their optimal values ​​are [values ​​to be filled in]. ;

[0184] The weighting coefficients representing the adversarial loss control the importance of local texture realism; the preferred value is [value missing]. ;

[0185] This represents the edge loss, calculated from the reconstruction step of the edge-aware decoder;

[0186] This represents the cross-entropy loss function, used to measure the consistency of the segmentation results;

[0187] The segmentation mask representing the prediction is obtained based on a pre-trained semantic segmentation network;

[0188] Represents a semantic segmentation mask with dimension . It is a One-Hot encoding mask. This indicates that the pixel does not belong to this class. This indicates that it belongs to this class.

[0189] In one implementation, the predicted segmentation mask Reconstructing images using a pre-trained semantic segmentation network The inference is performed, specifically using a segmentation model based on DeepLab or UNet. It outputs the class probability of each pixel, and then generates a One-Hot encoded mask based on the class probabilities, which is the predicted segmentation mask. .

[0190] It should be noted that, The term represents the reconstruction loss, and its role is to minimize pixel-level errors, ensuring the overall similarity between the reconstructed image and the real image. The edge loss characterization function aims to enhance the preservation of boundary details and improve visual clarity. The segmentation loss term serves to maintain the consistency of semantic information, ensuring that the reconstructed image does not affect high-level visual tasks. Adversarial loss is used to enhance the realism of local textures and prevent the reconstructed image from being too smooth. The combined design balances the influence of different losses through weight coefficients, enabling the network to optimize multiple objectives simultaneously in complex monitoring scenarios. This promotes feature sharing and regularization, improves the model's generalization ability and robustness, and maintains high performance, especially under uneven lighting and occlusion interference.

[0191] S506, Iterative Training of Data Augmentation Model

[0192] The iterative training process of the data augmentation model is based on a multi-task joint optimization objective, using the training dataset to perform end-to-end optimization of the constructed encoder-decoder network.

[0193] At the start of training, the model parameters are initialized, and the stored monitoring images and corresponding semantic segmentation masks are loaded as input.

[0194] In each iteration, forward propagation is first performed. The input image goes through illumination-invariant feature extraction, semantic segmentation mask generation, and dynamic occlusion synthesis steps. Then, the reconstruction image and the predicted segmentation mask are generated through a multi-scale feature-preserving encoder, attention-guided feature fusion, and edge-aware decoder reconstruction module.

[0195] Then, multi-task losses are calculated, including reconstruction loss, edge loss, segmentation loss and adversarial loss, which measure pixel-level similarity, edge detail preservation, semantic consistency and local texture realism, respectively.

[0196] The loss value is used to update the model parameters through the backpropagation algorithm, and the weights are adjusted using an optimizer (such as Adam) to minimize the total loss.

[0197] Iterative training continues, data is processed in batches and loss changes are monitored, and the training process also incorporates early stopping mechanisms and validation set evaluation to prevent overfitting.

[0198] The decision to stop iteration is based on performance metrics on the validation set. Training terminates when the total loss no longer decreases significantly over multiple consecutive periods or when the preset maximum number of iterations is reached, ensuring that the model achieves optimal generalization ability in complex monitoring scenarios.

[0199] S6, Enhanced Monitoring Data

[0200] Data augmentation for surveillance is implemented after the model training is completed. The trained data augmentation model is applied to new surveillance image data to improve image quality and semantic information integrity.

[0201] In the specific process, the newly acquired monitoring images are first loaded by the monitoring data storage module and then input into the enhancement system;

[0202] The system performs illumination-aware segmentation enhancement preprocessing, extracts illumination-invariant features and generates a semantic segmentation mask to eliminate illumination interference;

[0203] Then, the dynamic occlusion synthesis mechanism dynamically generates occlusion effects based on the segmentation mask, and embeds real occlusion objects into the background area through Poisson fusion to simulate temporary occlusion in real scenes;

[0204] The augmented image is then processed by a multi-scale feature-preserving encoder, attention-guided feature fusion, and edge-aware decoder reconstruction module of the data augmentation model to generate a high-quality reconstructed image that preserves clear boundaries and texture details while maintaining semantic consistency.

[0205] Ultimately, the enhanced images are output and stored for subsequent advanced vision tasks (such as object detection or behavior analysis), thereby significantly improving the robustness and accuracy of the monitoring system in complex environments.

[0206] In one embodiment, the feature selection performance of different attention mechanisms in complex scenes is demonstrated by visualizing feature response maps. Experiments compare the performance of no-attention mechanisms, spatial attention mechanisms, channel attention mechanisms, and the gated attention mechanism proposed in this invention. The experimental configuration simulates a complex monitoring scene containing multiple targets of different sizes. The horizontal and vertical axes of the feature response map represent spatial dimensions, and the color intensity represents the feature response strength, with darker red indicating a stronger response. It is clearly observed from the feature response maps that the feature response distribution of the no-attention mechanism is relatively uniform, with little difference between the background and target regions, lacking effective feature selection capabilities. The spatial attention mechanism enhances the feature response of the target region to some extent, but its suppression of background noise is insufficient. The channel attention mechanism further improves the feature response intensity of the target region, but its response at the target boundary is not precise enough. The gated attention mechanism of this invention exhibits the best feature selection performance, accurately enhancing the feature response of all target regions while maintaining clear response changes at the target boundary, and effectively suppressing background noise regions. Experimental results verify the technical advantages of the gated channel-spatial attention mechanism introduced in this invention. By dynamically adjusting the importance of features through multi-source information fusion, it can not only enhance the feature response of important channels and spatial locations, but also effectively suppress background noise and interference from irrelevant regions.

[0207] The above embodiments are provided to those skilled in the art to fully disclose and describe how the claimed implementations can be carried out and used, and are not intended to limit the scope of the disclosure herein. Modifications that will be obvious to those skilled in the art will be within the scope of the appended claims.

Claims

1. A method for storing and augmenting data in intelligent camera surveillance systems, characterized in that... Includes the following steps: S1. Capture diverse surveillance video streams using smart cameras, segment and label the captured raw videos, and obtain a training dataset; S2. Store the acquired monitoring data; S3. Enhance the preprocessing of monitoring data based on illumination perception segmentation, extract illumination-invariant features and generate semantic segmentation mask; S4. Establish a dynamic occlusion synthesis mechanism, dynamically generate occlusion effects based on the segmentation mask, and embed real occlusion objects into the background area through Poisson fusion to simulate temporary occlusion in real scenes. S5. Construct a data augmentation model, and process the data using the multi-scale feature-preserving encoder, attention-guided feature fusion, and edge-aware decoder reconstruction modules of the data augmentation model to generate high-quality reconstructed images. S6. Output and store the enhanced image for subsequent advanced vision tasks.

2. The intelligent camera monitoring data storage and data enhancement processing method according to claim 1, characterized in that: When acquiring the training dataset in S1, video data was continuously recorded under different time periods and weather conditions to ensure that the data covered complex lighting and occlusion scenarios.

3. The intelligent camera monitoring data storage and data enhancement processing method according to claim 1, characterized in that: In S2, the original monitoring images and labeled masks are stored in a distributed file system or cloud storage platform, and an index database is established to associate the image and mask files to ensure fast retrieval and access.

4. The intelligent camera monitoring data storage and data enhancement processing method according to claim 1, characterized in that: S3 specifically refers to: S31. Based on the original monitoring image, the dynamic range is compressed by logarithmic transformation, and Gaussian filtering is used to smooth the abrupt changes in illumination, thereby extracting illumination-invariant features. S32. Based on illumination-invariant features, a weighted Gaussian probability model is used to calculate the semantic category of each pixel. By combining the prior weights of the categories and the difference measure between the features and the category center, a semantic segmentation mask is generated.

5. The intelligent camera monitoring data storage and data enhancement processing method according to claim 1, characterized in that: S4 specifically refers to: S41. Based on the background semantic segmentation mask and the real occlusion image, the pose of the occlusion is adjusted by random transformation, and the transformed occlusion is seamlessly embedded into the background area using the Poisson fusion algorithm to generate a dynamic occlusion effect that conforms to the laws of physics. S42. Based on randomly generated occlusion masks and dynamic occlusion images, the occlusion areas are combined with the original images through weighted fusion to generate an occlusion enhancement image.

6. The intelligent camera monitoring data storage and data enhancement processing method according to claim 1, characterized in that: S5 specifically refers to: S51. A cross-scale residual aggregation module is adopted, which combines parallel downsampling paths and residual paths with a coordinate attention mechanism to achieve multi-scale feature preservation. S52. By calculating the channel attention-gated and spatial attention-gated weighted encoder features and deeply fusing them with the decoder features, combined with segmentation semantic information, the focus on important channels is enhanced. S53. By introducing segmentation boundary priors and multi-directional gradient loss, the reconstruction network is jointly optimized to improve the quality of edge detail restoration. S54. A patch-level discriminator structure is adopted, and adversarial training is carried out by randomly cropping local image patches to enhance the ability to preserve local texture features. S55. Based on reconstructed image, real image, edge loss, predictive segmentation mask, semantic segmentation mask and adversarial loss, a multi-task optimization objective is constructed by weighted combination of reconstruction loss, edge loss, segmentation loss and adversarial loss to balance pixel-level reconstruction quality, edge detail preservation, semantic consistency and local texture realism, so as to achieve end-to-end semantically consistent image enhancement. S56. Iteratively train the data augmentation model.

7. The intelligent camera monitoring data storage and data enhancement processing method according to claim 6, characterized in that: S51 specifically refers to: S511. Based on the input feature map and illumination-invariant features, downsampling is performed through convolution operation, and feature enhancement is performed by combining the scaled illumination-invariant features to generate a low-resolution feature map, so as to expand the receptive field and retain the robustness under low illumination. S512. Based on the segmentation prior of input feature map and coordinate attention weighting, nonlinear transformation is performed through dense connected blocks to generate high-resolution residual features, so as to preserve detailed information and enhance the local texture representation capability. S513. Based on downsampling features and upsampling residual features, multi-scale fusion features are generated by concatenating channels and fusing coordinate attention-weighted input features to simultaneously capture low-resolution semantic information and high-resolution detail information, thereby improving feature robustness. S514, The encoder's final layer output yields the encoder features. It integrates multi-scale downsampling features and residual features, and enhances spatial perception capabilities through a coordinate attention mechanism.

8. The intelligent camera monitoring data storage and data enhancement processing method according to claim 6, characterized in that: S52 specifically refers to: S521. Based on encoder features and semantic segmentation mask, channel information is extracted through global average pooling and global max pooling, and channel attention weights are generated by combining fully connected layers and Sigmoid activation. S522. Based on encoder features and illumination-invariant features, spatial information is extracted through average pooling and max pooling, and spatial attention weights are generated by combining convolution operations and Sigmoid activation. S523. Based on the encoder features weighted by channel attention gating and spatial attention gating, a fused feature is generated by adding it to the decoder features processed by depth-separable convolution.

9. The intelligent camera monitoring data storage and data enhancement processing method according to claim 6, characterized in that: S53 specifically refers to: S531. Based on fusion features, segmentation boundary features and multi-directional gradient features, an image reconstruction network is used to generate an edge-enhanced reconstructed image. S532. Based on the reconstructed image and the real image, the gradient difference is calculated through a multi-directional gradient operator and weighted by an edge weight mask to generate an edge loss.

10. The intelligent camera monitoring data storage and data enhancement processing method according to claim 6, characterized in that: S54 specifically refers to: S541. Based on the reconstructed image and semantic segmentation mask, local image patches are generated through random cropping to ensure that sampling is focused on the target region; S542. Based on real images and semantic segmentation masks, local image patches are generated through random cropping operations to ensure that sampling is focused on the target region; S543. Based on real and fake patches, a patch-level discriminator is used for discrimination, and the adversarial loss is calculated by combining the gradient penalty term.

Images