Method and system for enhancing video images of a security monitoring system

Abstract

The application discloses a security monitoring system video image enhancement method and system, belongs to the security image processing technical field, and aims to solve the technical problems that local overexposure, underexposure, noise, high concentration of air fine particulate matter and other mixed degradation in the security video picture are difficult to be uniformly and adaptively processed, the enhancement effect is dislocated with machine recognition targets, the enhanced model is difficult to adapt to the real noise characteristics of specific equipment after deployment and the like, pixel-level weighted fusion is performed on branch output features and three-dimensional lookup table vertex parameters, the enhancement result considers the machine recognition performance, the target camera shot noise and dark current noise coefficients are estimated by collecting calibration frames in the deployment stage, training samples are constructed to fine-tune the denoising branch and the lookup table parameter generation unit, the application can realize consistent repair on the mixed degradation picture, improve the downstream recognition accuracy, adapt to the real imaging link characteristics of different equipment, and guarantee long-term spatiotemporal consistency.

Description

Technical Field

[0001] This invention relates to the field of security image processing technology, specifically to a method and system for enhancing video images in security monitoring systems. Background Technology

[0002] In modern security monitoring systems, the quality of video images directly affects the accuracy of subsequent intelligent analysis tasks. Applications such as target detection, face recognition, and vehicle tracking are highly dependent on the clarity and structural integrity of the front-end video images. Since security cameras are usually in a continuous working state around the clock, the video images actually collected are easily interfered with by complex and changing environmental conditions, thus exposing several technical problems that urgently need to be solved.

[0003] Firstly, security cameras often face wide dynamic range scenarios such as backlit access control, tunnel entrances and exits, and direct headlights at night. Within the same frame, there are often severe local overexposure and local underexposure phenomena, which are often accompanied by compound degradation such as noise, rain and fog. Most existing image enhancement technologies are designed for single types of degradation and lack the ability to implement unified adaptive processing for mixed degradation with spatial variations within a single frame. This makes it difficult to balance the enhancement effect across different degradation areas, and inconsistent processing is likely to occur at degradation boundaries.

[0004] Secondly, most mainstream image enhancement algorithms currently aim to improve the subjective quality indicators of human vision. However, the core demand of security monitoring scenarios is often to improve the accuracy of downstream machine vision recognition tasks. Simply increasing the brightness and color saturation of the image can easily amplify high-frequency noise or erase texture structures that are important for machine recognition, thus causing a misalignment between human perception and machine recognition results.

[0005] Third, most existing video image enhancement models are trained offline based on synthetic degradation data under ideal laboratory conditions, without considering the noise characteristics of the actual imaging links of the camera sensors in specific deployments. Due to the differences in image signal processing link parameters of different brands and batches of cameras, and the spatiotemporal uneven distribution of real noise, enhancement models with fixed parameters often experience a sharp drop in performance after being deployed to real devices, and may even produce flickering and breathing effects between video frames, making it difficult to maintain long-term spatiotemporal consistency.

[0006] Currently, the closest technical solution to this invention in the industry is a low-light image enhancement method based on deep convolutional neural networks, such as enhancement networks based on image decomposition theory or lightweight networks based on pixel-level enhancement curve estimation. The general implementation of this type of method is as follows: acquire low-light video frames to be processed, extract features from the input image through an encoder network to obtain multi-scale feature maps, predict pixel-level or block-level enhancement curve parameters through an enhancement parameter estimation module, apply the predicted parameters to the original image to obtain the enhanced output, and optimize the network parameters using the pixel-level reconstruction error and perceptual loss between the enhanced image and the reference ground image as training targets. This type of method does not introduce feedback signals from downstream recognition tasks during the training and inference stages, nor does it perform on-site calibration for the imaging link noise characteristics of the specific deployed cameras. Its enhancement network usually adopts a single-branch structure and uses a uniform processing strategy for the entire image, which is difficult to address the aforementioned three technical problems. Therefore, we propose a video image enhancement method and system for security monitoring systems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for enhancing video images in a security monitoring system.

[0008] To address the shortcomings of the existing technologies, the technical problems to be solved by this invention include the following three aspects: First, how to achieve unified spatial adaptive enhancement processing in wide dynamic range scenarios where a single frame simultaneously exhibits local overexposure, local underexposure, and various mixed degradations such as noise, rain, and fog, overcoming the problem that a single-branch enhancement network cannot adequately handle complex degradations; second, how to ensure that the enhanced image effectively improves the accuracy of downstream machine recognition tasks, eliminating the target misalignment between enhancement and recognition effects caused by using human visual quality indicators as the sole optimization objective; and third, how to ensure that the enhancement model obtained through offline training can automatically adapt to the noise characteristics of the actual imaging link of a specific security camera model after deployment, overcoming the performance drop caused by domain differences and inter-frame flickering problems.

[0009] To address the aforementioned technical problems, this invention provides a method for enhancing video images in a security monitoring system. The method includes: acquiring a video frame to be processed output from a security camera; extracting shared features from the video frame using a feature extraction backbone network; estimating pixel-level degradation type and intensity of the shared features using a degradation-aware routing sub-network to obtain a degradation vector containing four components: overexposure degradation intensity, underexposure degradation intensity, noise degradation intensity, and high concentration of fine particulate matter degradation intensity at the same pixel location; inputting the degradation vector into a routing weight generation unit to obtain normalized activation weights for the exposure compensation branch, denoising branch, dehazing branch, and identity mapping branch corresponding to the pixel location; inputting the shared features into the four enhancement branches for parallel processing; performing pixel-level weighted summation of the output features of each branch according to the normalized activation weights to obtain fused features; simultaneously, performing weighted summation of the pre-learned 3D lookup table vertex parameters for each branch according to the normalized activation weights to obtain the 3D lookup table vertex parameters corresponding to the pixel location; and performing hue and contrast mapping on the fused features using trilinear interpolation to obtain the final enhanced image.

[0010] Building upon this foundation, the present invention further provides a task-driven closed-loop joint optimization mechanism. During the training phase, the enhanced video frames are input into the downstream task network. The task-related loss is calculated based on the error between the task prediction result and the task label. The gradient of this loss is then backpropagated to the degradation-aware routing sub-network and the learnable parameters of each enhancement branch, directly linking the training objective of the enhancement network with the performance of the downstream recognition task. The present invention also provides an adaptive calibration method for deployment scenarios based on the inversion of imaging link noise parameters from an image signal processor. After the security camera is deployed, the shot noise coefficient and dark current noise coefficient corresponding to the device are estimated by collecting calibration frames. Based on this, training samples are constructed to fine-tune the denoising branch and the 3D lookup table parameter generation unit with small samples, enabling the enhancement model to automatically adapt to the real noise characteristics of the specific deployment device.

[0011] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0012] 1. This invention combines pixel-level degradation type and intensity blind perception evaluation with a multi-branch weighted fusion mechanism to perform differentiated processing on mixed degradation with uneven spatial distribution within a single frame, avoiding local area processing imbalance and achieving consistent restoration of mixed degradation within the same frame;

[0013] 2. This invention introduces task-related loss into the training objective of the augmentation network and performs gradient backpropagation, so that the augmentation result adaptively highlights the feature frequencies that are beneficial to machine recognition, thereby effectively improving the accuracy of downstream recognition tasks while ensuring the visual quality of the image.

[0014] 3. This invention enables the enhancement model to automatically adapt to the real imaging link characteristics of different devices by estimating the on-site noise parameters of the specific deployed cameras and fine-tuning them online with small samples. This improves the long-term spatiotemporal consistency of the enhancement effect under complex weather and extreme lighting conditions and reduces inter-frame flicker. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall video enhancement main process in an embodiment of the present invention;

[0016] Figure 2 This is a schematic diagram of the task-driven joint optimization process during the training phase in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram of the online noise calibration and fine-tuning process after device deployment in an embodiment of the present invention. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0019] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0021] In this embodiment, the video stream collected by a certain type of security camera is used as the processing object. Its video input resolution is 1,920 x 1,080 pixels, and it is input frame by frame into the enhancement network of the present invention for processing.

[0022] First, the security camera outputs the original video frame through its internal image signal processing link. The video frame is then sent to the feature extraction backbone network. The feature extraction backbone network uses a multi-layer convolutional structure to perform downsampling feature extraction on the input video frame layer by layer, and outputs a shared feature map with multiple scales, which can be used by the subsequent degradation perception routing module and the multi-branch enhancement sub-network to realize the reuse of feature calculation and reduce the overall computational overhead.

[0023] Secondly, the shared features are fed into a degradation-aware routing subnetwork, which extracts degradation type and intensity information for each pixel location to obtain a degradation vector:

[0024] ;

[0025] in, Indicates the location of pixels or local regions in an image. , , , These represent the intensity estimates of four types of degradation at this location: overexposure, underexposure, noise, and high concentration of fine particulate matter in the air. The values ​​are all continuous values ​​between 0 and 1. The closer the value is to 1, the more significant the degradation is at this location. The degradation-aware routing subnetwork can be implemented using a lightweight convolutional structure. Its training can be combined with supervised learning using synthetic data labeled with degradation types.

[0026] Subsequently, the degradation vector is input into the routing weight generation unit, which consists of a small multilayer sensing mechanism, and outputs the position relative to the first... Normalized activation weights for each enhancement branch:

[0027] ;

[0028] in, Indicates the first Each branch corresponds to a learnable nonlinear mapping function. This represents the total number of enhanced branches, in this embodiment. The value is 4, corresponding to the exposure compensation branch, noise reduction branch, dehazing branch, and identity mapping branch, respectively. Indicates the first Each branch is in position The normalized activation weights at the point satisfy that the sum of the weights of the four branches is always equal to 1. The above exponential normalization process enables the routing weight generation unit to adaptively concentrate processing resources on the enhancement branches that match the current degradation type, while assigning lower weights to the other branches, based on the distribution of the degradation vector.

[0029] Next, shared features They are fed into four parallel branch networks respectively. The processing involves four branches: an exposure compensation branch, which adjusts brightness and contrast in overexposed and underexposed areas; a denoising branch, which smooths features in noisy areas; a dehazing branch, which enhances contrast and restores color in areas with high concentrations of fine particulate matter; and an identity mapping branch, which preserves the original features of areas with low degradation levels, avoiding unnecessary processing of already high-quality image areas. All four branches can be implemented using a lightweight network structure with several layers of dynamic convolutions. The kernel parameters of the dynamic convolutions can be adaptively generated based on the input content, reducing computational overhead while maintaining processing effectiveness. The output features of the four branches are then pixel-weighted and fused according to routing weights to obtain the fused features.

[0030] ;

[0031] in, Indicates the first Branches to positions The processing results of the input features, The output features obtained after fusion are represented by the weights dynamically generated according to the pixel position. This fusion process can achieve spatial adaptive effect of using different processing strategies in different areas within the same image, avoiding the problem of inconsistent degradation boundaries caused by uniform processing of a single branch.

[0032] While obtaining the fusion features, this embodiment further dynamically generates the vertex parameters of the 3D lookup table corresponding to the position based on the weights of each branch:

[0033] ;

[0034] in, Indicates prior to the first The set of basic vertex parameters for the 3D lookup table obtained from each branch learning is jointly optimized and determined with the network parameters of each branch during the network training phase. Indicates position The actual lookup table parameters used after weighted merging are as follows: To look up table parameters, fused features are processed using trilinear interpolation units. Perform hue and contrast mapping to obtain the final enhanced output at that location. Since the vertex parameters of the 3D lookup table are also dynamically generated based on the pixel-level routing weights, the tone mapping process also has spatial adaptive capabilities, which can adopt tone adjustment strategies that match different degradation regions, further improving the enhanced consistency in composite degradation scenarios.

[0035] After the enhanced image output is sent, one path leads to the display and storage end for manual review, while the other path feeds it into the task-driven closed-loop joint optimization module during the training phase. This module enhances the image output by the network. The data are fed into the reconstruction loss calculation unit and the downstream task network, respectively. The downstream task network can be one of the pre-trained face recognition network, license plate recognition network, or abnormal behavior detection network. Its parameters can be frozen in the early stage of joint training, or they can be fine-tuned synchronously with a small learning rate. After the downstream task network outputs the task prediction result, the task-related loss is calculated based on the error between the prediction result and the task label, and the overall joint loss is also calculated.

[0036] ;

[0037] in, This represents the pixel-level reconstruction loss between the enhanced image and the ground truth reference image. This represents the perceptual loss calculated based on the pre-trained feature extraction network. This represents the task-related loss between the downstream task network output and the task label. Specifically, it can be one of the following: bounding box regression loss, classification cross-entropy loss, and recognition feature contrast loss. , , These are the weighting coefficients corresponding to the three losses, used to balance the relationship between visual quality and task performance. Their specific values ​​can be preset according to the relative importance of visual display requirements and recognition accuracy requirements in the actual deployment scenario.

[0038] By using the chain rule, the gradient of the task-related loss is backpropagated to the learnable parameters of the enhancement network:

[0039] ;

[0040] in, This represents the set of learnable parameters for the degraded sensing routing subnetwork, the routing weight generation unit, and the four enhancement branches in the augmentation network. The gradient obtained from the above equation is used together with the gradients generated by the reconstruction loss and the sensing loss to update... Gradient descent optimizers are used based on total loss. The parameters of the augmentation network are iteratively updated until the task-related loss and the visual reconstruction loss converge simultaneously, thus completing the task-driven closed-loop joint training. Since this training process directly guides the update of the augmentation network parameters based on the performance feedback of the downstream recognition task, the augmentation network can adaptively retain and highlight high-frequency texture and structural features that are beneficial to machine recognition while maintaining the visual effect of the image. This fundamentally alleviates the target misalignment problem caused by the previous optimization target that only used human visual indicators.

[0041] During the equipment deployment phase, this embodiment further performs adaptive calibration of the deployment scenario based on image signal processor parameter inversion. After the security camera is deployed, it is controlled to acquire data. A set of calibration frames under different signal intensities, including a set of dark-field frames and multiple sets of flat-field frames under different exposure conditions, were used. The pixel mean of each set of calibration frames was calculated. With pixel variance ,in Based on a pre-defined noise model where sensor noise and signal strength have a linear relationship:

[0042] ;

[0043] in, Indicates pixel signal strength. This represents the noise variance corresponding to this signal strength. This represents the proportionality coefficient associated with shot noise, reflecting the physical property that noise caused by the randomness of photons arriving at the sensor is approximately proportional to the signal strength. This represents the bias coefficient related to dark current and readout noise. This coefficient reflects the noise component generated by thermal excitation of the sensor under no light or low illumination conditions, which is basically independent of the signal strength.

[0044] For the multiple sets of mean-variance data pairs collected, the optimization objective is to minimize the sum of squared errors between the pixel variance of each calibration frame and the predicted variance calculated based on the above noise model, and then solve for the true noise parameters of the camera.

[0045] ;

[0046] in, , These represent the estimated values ​​of the shot noise figure and dark current noise figure obtained from the solution, respectively. The solution process for these parameters can be achieved using the least squares method, which has the advantages of low computational load and can be completed in real time at the equipment deployment site.

[0047] Using the estimated , For a small number of clear reference images collected on-site, corresponding degraded images are synthesized according to the above noise model. Training sample pairs corresponding to the clear reference images and the degraded images that conform to the noise characteristics of the camera's real imaging link are constructed. Using the training sample pairs as training data, small sample gradient fine-tuning is performed on the denoising branch and the vertex parameter generation unit of the 3D lookup table in the enhancement network that has completed offline training. During the fine-tuning process, the parameters of the degraded perception routing subnetwork, the routing weight generation unit, and the exposure compensation branch, defogging branch, and identity mapping branch are frozen, and only the local parameters directly related to the noise characteristics are updated to obtain the final enhancement model adapted to the currently deployed equipment. This calibration process can be repeated periodically during the equipment's operating cycle to adapt to the noise characteristic drift caused by seasonal and weather changes in the on-site environment, ensure the long-term spatiotemporal consistency of the enhancement effect, and reduce the flickering and breathing effect between video frames.

[0048] The aforementioned three components—spatial adaptive partitioning enhancement, task-driven closed-loop joint optimization, and deployment scenario adaptive calibration—work together on a unified augmentation network backbone. During the offline training phase, the spatial adaptive partitioning enhancement mechanism and the task-driven closed-loop joint optimization mechanism work together to enable the augmentation network to learn spatial adaptive processing capabilities while also considering downstream recognition performance. During the device deployment phase, the deployment scenario adaptive calibration mechanism fine-tunes local parameters for specific devices based on the trained model, enabling the augmentation network to ultimately possess comprehensive performance that balances hybrid degradation processing, machine recognition friendliness, and device adaptability.

[0049] In another embodiment of the present invention, the generation and mapping method of the vertex parameters of the three-dimensional lookup table can be replaced by the bilateral mesh parameters. That is, the basic parameters of the bilateral mesh are learned in advance for each enhancement branch, and the basic parameters of the bilateral mesh of each branch are weighted and summed according to the normalized activation weight to obtain the bilateral mesh parameters corresponding to the position. The fusion features are then mapped to tone and contrast through bilateral mesh interpolation, which can also achieve spatial adaptive tone adjustment effect while ensuring real-time processing.

[0050] In another embodiment of the present invention, the task-related loss in the task-driven closed-loop joint optimization can be replaced by the feature-level knowledge distillation loss. That is, the features generated by the constrained enhanced image in the intermediate layer of the downstream task network are consistent with the features generated by the ideal clear image in the intermediate layer of the same network. The gradient backpropagation of the feature alignment loss indirectly guides the update of the enhanced network parameters, which can also achieve the optimization of the enhancement effect for machine recognition tasks.

[0051] It should be noted that the specific network layer number, channel number, branch number, and resolution parameters given in the above embodiments are merely examples to facilitate the illustration of the technical solution of the present invention. Those skilled in the art can adjust the above parameters according to the actual needs of specific security monitoring scenarios based on the technical solution of the present invention. Any technical solution implemented using the same or similar technical means as the present invention should fall within the protection scope of the present invention.

[0052] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. Methods for enhancing video images in security monitoring systems, including: The system acquires video frames from security cameras, extracts features from these frames using a feature extraction backbone network to obtain shared features, and further includes: estimating pixel-level degradation type and intensity using a degradation-aware routing sub-network to obtain a degradation vector. This degradation vector contains four components representing the degradation intensity at the same pixel location: overexposure degradation intensity, underexposure degradation intensity, noise degradation intensity, and degradation intensity due to high concentration of fine particulate matter in the air. The degradation vector is then input into a routing weight generation unit to obtain normalized activation weights for the pixel location corresponding to the exposure compensation branch, denoising branch, dehazing branch, and identity mapping branch, respectively. The sum of these four normalized activation weights... If the value is 1, the shared features are input into the exposure compensation branch, denoising branch, dehazing branch, and identity mapping branch for parallel processing to obtain four branch output features. The four branch output features are then pixel-level weighted summed according to the normalized activation weights to obtain the fused features. The basic vertex parameters of the 3D lookup table, which are pre-learned for the exposure compensation branch, denoising branch, dehazing branch, and identity mapping branch, are then weighted summed according to the normalized activation weights to obtain the 3D lookup table vertex parameters corresponding to the pixel position. Using the 3D lookup table vertex parameters as lookup table parameters, the fused features are mapped to tone and contrast through trilinear interpolation to obtain the enhanced video frame.

2. The method for enhancing video images in a security monitoring system according to claim 1, characterized in that: The values ​​of the overexposure degradation intensity, the underexposure degradation intensity, the noise degradation intensity, and the degradation intensity due to high concentration of fine particulate matter in the air are all between 0 and 1. The normalized activation weight is obtained by performing a learnable nonlinear mapping on the degradation vector and then performing exponential normalization on the mapping result.

3. The method for enhancing video images in a security monitoring system according to claim 1, characterized in that, Also includes: During the training phase, the enhanced video frames are input into the downstream task network to obtain task prediction results. Based on the error between the task prediction results and the task labels, the task-related loss is calculated. The gradient of the task-related loss is backpropagated along the enhanced video frames to the learnable parameters of the degradation-aware routing subnetwork, the routing weight generation unit, and the exposure compensation branch, denoising branch, dehazing branch, and identity mapping branch, and the learnable parameters are updated.

4. The method for enhancing video images in a security monitoring system according to claim 3, characterized in that: Calculate the pixel-level reconstruction loss between the enhanced video frame and the reference ground truth image, and the perceptual loss calculated based on the pre-trained feature extraction network. Then, sum the reconstruction loss, the perceptual loss, and the task-related loss according to preset weighting coefficients to obtain the joint loss. Use a gradient descent optimizer to iteratively update the learnable parameters based on the joint loss.

5. The method for enhancing video images in a security monitoring system according to claim 1, characterized in that, Also includes: After the security camera is deployed, it is controlled to collect multiple sets of calibration frames under different signal strengths. The calibration frames include dark field frames and multiple flat field frames under different exposure conditions. The pixel mean and pixel variance of each set of calibration frames are calculated. Based on a preset noise model in which the pixel signal strength and noise variance are linearly related, the pixel mean and pixel variance of multiple sets are fitted with parameters to obtain the shot noise coefficient and dark current noise coefficient corresponding to the security camera.

6. The method for enhancing video images in a security monitoring system according to claim 5, characterized in that: The parameter fitting employs a parameter estimation method with the goal of minimizing the sum of squared errors. The goal is to minimize the sum of squared errors between the pixel variance of each calibration frame and the variance estimate calculated based on the preset noise model. This is used to solve for the shot noise coefficient and the dark current noise coefficient. Using these coefficients, a degraded image is synthesized from a clear reference image acquired on-site according to the preset noise model, forming training sample pairs corresponding to the degraded image and the clear reference image. These training sample pairs are then used as training data to fine-tune the parameters of the denoising branch and the vertex parameter generation unit of the 3D lookup table. During this fine-tuning process, the parameters of the degradation-aware routing subnetwork, the exposure compensation branch, the dehazing branch, and the identity mapping branch are frozen.

7. The method for enhancing video images in a security monitoring system according to claim 1, characterized in that: The vertex parameters of the 3D lookup table are replaced with bilateral grid parameters, and the fused features are mapped to tone and contrast through bilateral grid interpolation to obtain the enhanced video frame.

8. A video image enhancement system for a security monitoring system, comprising: A video acquisition interface is used to acquire video frames to be processed output by a security camera. A feature extraction backbone network is used to extract features from the video frames to obtain shared features. The interface further includes: a degradation-aware routing module connected to the feature extraction backbone network, used to estimate the pixel-level degradation type and intensity of the shared features to obtain a degradation vector, and to generate normalized activation weights corresponding to the exposure compensation branch, denoising branch, dehazing branch, and identity mapping branch based on the degradation vector; and a multi-branch enhancement sub-network connected to the feature extraction backbone network and the degradation-aware routing module, including parallel exposure compensation branches, denoising branches, and identity mapping branches. The fog branch and the identity mapping branch are used to process the shared features to obtain branch output features, and to perform pixel-level weighted summation of the branch output features according to the normalized activation weights to obtain fused features. The lookup table generation and mapping unit is connected to the degradation-aware routing module and the multi-branch enhancement sub-network. It is used to perform weighted summation of the basic vertex parameters of the three-dimensional lookup table pre-learned for the exposure compensation branch, the noise reduction branch, the defogging branch and the identity mapping branch according to the normalized activation weights to obtain the vertex parameters of the three-dimensional lookup table. Then, the fused features are mapped to tone and contrast through trilinear interpolation to obtain the enhanced video frame.

9. The video image enhancement system for a security monitoring system according to claim 8, characterized in that: It also includes a downstream task network and a joint loss calculation unit. The downstream task network is connected to the lookup table generation and mapping unit and is used to process the enhanced video frames during the training phase to obtain task prediction results. The joint loss calculation unit is used to calculate task-related loss based on the error between the task prediction results and the task labels, and backpropagate the gradient of the task-related loss to the degradation-aware routing module and the learnable parameters of the multi-branch enhancement sub-network.

10. The video image enhancement system for a security monitoring system according to claim 8, characterized in that: It also includes an online calibration module, which is connected to the security camera, the denoising branch in the multi-branch enhancement sub-network, and the lookup table generation and mapping unit. After the security camera is deployed, it receives the calibration frames collected by the security camera, estimates the shot noise figure and dark current noise figure corresponding to the security camera, and fine-tunes the parameters of the denoising branch and the lookup table generation and mapping unit using training samples constructed from the shot noise figure and the dark current noise figure.

Images