Short video key region enhancement method and system based on visual attention mechanism

Abstract

The present application belongs to the field of computer vision and multimedia intelligent processing, and specifically relates to a short video key region enhancement method and system based on a visual attention mechanism. In view of the defects of the existing short video enhancement technology, such as the global unified enhancement strategy, the neglect of user visual attention, the unnatural enhancement effect and the poor real-time performance, a full-process enhancement system covering video decoding preprocessing, multi-modal key region detection, spatio-temporal joint visual attention modeling, hierarchical adaptive enhancement and video encoding output is constructed. The system combines multiple technologies such as face detection and text recognition to realize accurate positioning of multi-dimensional key regions, designs a spatio-temporal double-branch attention network to generate a pixel-level enhancement weight map, differentially processes key and background regions through a hierarchical adaptive algorithm, and introduces an inter-frame consistency constraint to eliminate flicker.

Description

Technical Field

[0001] This invention relates to the field of short video processing technology, specifically to a method and system for enhancing key regions in short videos based on visual attention mechanisms. Background Technology

[0002] As short videos become the core medium for information dissemination and commercial promotion, original videos in complex scenarios suffer from problems such as unclear key areas and loss of details, which seriously affect the viewing experience and information transmission efficiency.

[0003] Existing short video enhancement technologies suffer from four major flaws: First, the use of a globally unified enhancement strategy easily leads to insufficient enhancement of key areas, over-enhancement of the background, and color distortion. Second, the key area recognition capability is limited, only able to recognize faces, with low accuracy in detecting text, products, and small, blurry targets. Third, the lack of spatiotemporal attention modeling results in frame-to-frame flickering due to independent processing of each frame, and it is also impossible to continuously track dynamic key areas. Fourth, it is difficult to balance real-time performance with enhancement effects, with high-precision models exhibiting slow inference and lightweight models performing poorly.

[0004] Existing patents have also failed to solve the above problems, and cannot simultaneously meet the needs of precise enhancement of key areas, spatiotemporal consistency and real-time processing. There is an urgent need for a new short video enhancement solution based on visual attention mechanisms. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for enhancing key regions in short videos based on visual attention mechanisms, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a short video key region enhancement system based on visual attention mechanism, comprising:

[0007] The video preprocessing module is used to decode, extract frames and convert formats of the input short video to generate a continuous sequence of raw video frames, and perform noise reduction and normalization preprocessing.

[0008] The key region detection module integrates face detection, text recognition, object detection and motion estimation technologies to perform multi-dimensional analysis on each original video frame, identify all key regions and generate an initial region mask.

[0009] The spatiotemporal attention modeling module is used to construct a spatiotemporal dual-branch visual attention network, which combines the key region distribution in the spatial dimension with the inter-frame motion information in the temporal dimension to generate a pixel-level spatiotemporal joint attention weight map.

[0010] The hierarchical adaptive enhancement module is used to divide video frames into core key regions, secondary key regions and background regions based on the spatiotemporal joint attention weight map, and to apply different enhancement algorithms and parameters to different regions to achieve hierarchical adaptive enhancement.

[0011] The video post-processing module is used to perform inter-frame consistency constraints, color correction, and format conversion on the enhanced video frame sequence, and finally encode and output the enhanced short video.

[0012] Preferably, the video preprocessing module is specifically used to decode the input short video, extract the original video frame sequence in YUV format, perform Gaussian filtering noise reduction on the luminance component, and normalize the pixel values ​​to the [0,1] range.

[0013] Preferably, the key region types identified by the key region detection module include face regions, text regions, product regions, and moving object regions. The logic for generating the initial region mask is as follows: corresponding region masks are generated by using face detection algorithm, text recognition algorithm, object detection algorithm, and motion estimation algorithm respectively, and the initial key region mask is obtained by performing a logical OR operation on all region masks.

[0014] Preferably, the spatiotemporal bi-branch visual attention network includes a spatial attention branch and a temporal attention branch, and the logic for generating the spatiotemporal joint attention weight map is as follows:

[0015] The spatial attention branch takes the original video frame of the current frame and the initial key region mask as input, and extracts multi-scale spatial features through a convolutional neural network to generate a spatial attention weight map.

[0016] The temporal attention branch takes the current frame and the previous N original video frames as input, extracts inter-frame motion features through a 3D convolutional neural network, and generates a temporal attention weight map.

[0017] A learnable weighted fusion strategy is adopted to fuse the spatial attention weight map and the temporal attention weight map to generate a spatiotemporal joint attention weight map, which is then Gaussian smoothed.

[0018] Preferably, the fusion formula for the spatiotemporal joint attention weight graph is:

[0019] in, This is a spatial attention weight map. For time-based attention weights, and For learnable fusion weight coefficients, satisfying . The preferred layered adaptive enhancement module divides the video frame into a core critical region, a secondary critical region, and a background region. For the core critical region, it employs adaptive histogram equalization and unsharpening mask enhancement based on guided filtering. For the secondary critical region, it uses linear contrast stretching enhancement. For the background region, it uses gamma correction enhancement.

[0020] The preferred formula for pixel-level brightness enhancement is:

[0021]

[0022] in, This is the original pixel brightness value. To enhance the brightness value of the subsequent pixels, This is the brightness enhancement factor. The spatiotemporal joint attention weight value is the corresponding pixel.

[0023] Preferably, the inter-frame consistency constraint logic of the video post-processing module is as follows: calculate the average difference in brightness between adjacent frames after enhancement; if the difference exceeds a preset threshold, then linearly adjust the brightness of the next frame to eliminate inter-frame brightness jumps and flickering.

[0024] Preferably, the key region detection module uses MTCNN or RetinaFace for face detection, CRNN or PaddleOCR for text recognition, YOLOv5 or YOLOv8 for object detection, and optical flow or frame difference for motion estimation.

[0025] An application method for a short video key region enhancement system based on visual attention mechanism includes the following steps:

[0026] S1, Video Preprocessing: Decode, extract, denoise, and normalize the input short video to generate the original video frame sequence;

[0027] S2, Key Region Detection: Integrates multi-dimensional detection technology to identify all key regions in the video frame and generate an initial region mask;

[0028] S3, Spatiotemporal attention modeling: Pixel-level spatiotemporal joint attention weight map is generated through a spatiotemporal bi-branch visual attention network;

[0029] S4, Layered Adaptive Enhancement: Differentiated enhancement is applied to different regions based on the attention weight map to achieve precise enhancement of key regions;

[0030] S5, video post-processing: performs inter-frame consistency constraints, color correction and encoding, and outputs enhanced short videos.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. Accurate identification of key areas in multiple dimensions: It integrates multi-dimensional detection technologies for faces, text, products, and moving objects, and can comprehensively identify various key information in short videos, solving the problem that traditional methods can only process a single key area;

[0033] 2. By using a spatiotemporal dual-branch network, which simultaneously considers the distribution of key regions in the spatial dimension and the inter-frame motion information in the temporal dimension, pixel-level enhancement weights are generated, achieving continuous tracking enhancement of dynamic key regions.

[0034] 3. Differentiated enhancement is applied to different regions based on attention weights, with key enhancements to details in core regions and a natural appearance in background regions, avoiding over-enhancement and distortion issues caused by global enhancement;

[0035] By introducing an inter-frame brightness adjustment mechanism, the inter-frame flickering phenomenon caused by independent processing of single frames is effectively eliminated, significantly improving the smoothness of the video. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 This invention provides a technical solution: a short video key region enhancement system based on visual attention mechanism, comprising:

[0041] The video preprocessing module is used to decode, extract frames and convert formats of the input short video to generate a continuous sequence of raw video frames, and perform noise reduction and normalization preprocessing.

[0042] The key region detection module integrates face detection, text recognition, object detection and motion estimation technologies to perform multi-dimensional analysis on each original video frame, identify all key regions and generate an initial region mask.

[0043] The spatiotemporal attention modeling module is used to construct a spatiotemporal dual-branch visual attention network, which combines the key region distribution in the spatial dimension with the inter-frame motion information in the temporal dimension to generate a pixel-level spatiotemporal joint attention weight map.

[0044] The hierarchical adaptive enhancement module is used to divide video frames into core key regions, secondary key regions and background regions based on the spatiotemporal joint attention weight map, and to apply different enhancement algorithms and parameters to different regions to achieve hierarchical adaptive enhancement.

[0045] The video post-processing module is used to perform inter-frame consistency constraints, color correction, and format conversion on the enhanced video frame sequence, and finally encode and output the enhanced short video.

[0046] The video preprocessing module is specifically used to decode the input short video, extract the original video frame sequence in YUV format, perform Gaussian filtering on the luminance component for noise reduction, and normalize the pixel values ​​to the [0,1] range.

[0047] The key region detection module identifies key region types including face regions, text regions, product regions, and moving object regions. The logic for generating the initial region mask is as follows: the face detection algorithm, text recognition algorithm, object detection algorithm, and motion estimation algorithm are used to generate corresponding region masks, and all region masks are logically ORed to obtain the initial key region mask.

[0048] The spatiotemporal bi-branch visual attention network includes a spatial attention branch and a temporal attention branch. The logic for generating the spatiotemporal joint attention weight map is as follows:

[0049] The spatial attention branch takes the original video frame of the current frame and the initial key region mask as input, and extracts multi-scale spatial features through a convolutional neural network to generate a spatial attention weight map.

[0050] The temporal attention branch takes the current frame and the previous N original video frames as input, extracts inter-frame motion features through a 3D convolutional neural network, and generates a temporal attention weight map.

[0051] A learnable weighted fusion strategy is adopted to fuse the spatial attention weight map and the temporal attention weight map to generate a spatiotemporal joint attention weight map, which is then Gaussian smoothed.

[0052] The fusion formula for the spatiotemporal joint attention weight graph is:

[0053]

[0054] in, This is a spatial attention weight map. For time-based attention weights, and For learnable fusion weight coefficients, satisfying .

[0055] The layered adaptive enhancement module divides video frames into core key regions ( Secondary key areas ( ) and background area ( For the core critical areas, adaptive histogram equalization and unsharpened mask enhancement based on guided filtering are used; for the secondary critical areas, linear contrast stretching enhancement is used; and for the background areas, gamma correction enhancement is used.

[0056] The formula for pixel-level brightness enhancement is:

[0057]

[0058] in, This is the original pixel brightness value. To enhance the brightness value of the subsequent pixels, This is the brightness enhancement factor. The spatiotemporal joint attention weight value is the corresponding pixel.

[0059] The inter-frame consistency constraint logic of the video post-processing module is as follows: calculate the average difference in brightness between adjacent frames after enhancement. If the difference exceeds a preset threshold, then the brightness of the next frame is linearly adjusted to eliminate inter-frame brightness jumps and flickering.

[0060] The key region detection module uses MTCNN or RetinaFace for face detection, CRNN or PaddleOCR for text recognition, YOLOv5 or YOLOv8 for object detection, and optical flow or frame difference for motion estimation.

[0061] An application method for a short video key region enhancement system based on visual attention mechanism includes the following steps:

[0062] S1, Video Preprocessing: Decode, extract, denoise, and normalize the input short video to generate the original video frame sequence;

[0063] S2, Key Region Detection: Integrates multi-dimensional detection technology to identify all key regions in the video frame and generate an initial region mask;

[0064] S3, Spatiotemporal attention modeling: Pixel-level spatiotemporal joint attention weight map is generated through a spatiotemporal bi-branch visual attention network;

[0065] S4, Layered Adaptive Enhancement: Differentiated enhancement is applied to different regions based on the attention weight map to achieve precise enhancement of key regions;

[0066] S5, video post-processing: performs inter-frame consistency constraints, color correction and encoding, and outputs enhanced short videos.

[0067] The purpose of this invention is to provide a method and system for enhancing key regions of short videos based on visual attention mechanisms. By using multimodal key region detection and spatiotemporal attention modeling, this invention solves the problems of non-targeted global enhancement, incomplete key region recognition, inter-frame flickering, and poor real-time performance in existing technologies. It achieves accurate adaptive enhancement of key regions in short videos, thereby improving the information transmission efficiency of video content and the user viewing experience.

[0068] Specific module descriptions:

[0069] Video preprocessing module

[0070] The input short video is decoded to extract the original video frame sequence in YUV format, with a frame rate of 25fps or 30fps and supported resolutions of 720p, 1080p, or 4K. The luminance component Y of each frame is subjected to 3×3 Gaussian filtering to remove Gaussian noise and salt-and-pepper noise generated during shooting. The pixel values ​​of the denoised video frames are normalized to the [0,1] interval to prepare for subsequent deep learning model processing.

[0071] Key area detection module

[0072] The key regions to be identified include face regions, text regions, product regions, moving object regions, and user interaction regions. The logic for generating the initial region mask is as follows:

[0073] 1) Use MTCNN or RetinaFace face detection algorithms to detect all face regions in the video frame, generate face region masks, and set the confidence threshold to 0.7 to filter out false detections with low confidence.

[0074] 2) Use CRNN or PaddleOCR text recognition algorithms to detect all text regions in the video frame, generate text region masks, and set the confidence threshold to 0.6 to ensure complete detection of text regions;

[0075] 3) Use YOLOv5 or YOLOv8 object detection algorithms to detect preset product categories (such as mobile phones, cosmetics, food, etc.) in video frames, generate product area masks, and set the confidence threshold to 0.65;

[0076] 4) Motion estimation is performed using optical flow or frame difference methods. Motion vectors between adjacent frames are calculated, and regions of moving objects whose motion amplitude exceeds a preset threshold are identified to generate motion region masks.

[0077] 5) Perform a logical OR operation on all the above region masks to obtain the initial key region mask. In the mask, a pixel value of 1 represents the key region and 0 represents the background region.

[0078] Spatiotemporal attention modeling module

[0079] The constructed spatiotemporal bi-branch visual attention network includes a spatial attention branch and a temporal attention branch. The logic for generating the spatiotemporal joint attention weight map is as follows:

[0080] 1) Spatial Attention Branch: Taking the original video frame of the current frame and the initial key region mask as input, a three-layer convolutional neural network is used to extract multi-scale spatial features, which are then combined with the initial region mask to generate a spatial attention weight map. ,in The height of the video frame. The width of the video frame is used, and the weight value ranges from [0,1]. The larger the value, the more important the pixel is.

[0081] 2) Temporal attention branch: This branch considers the current frame and the previous frame... frame( Using the original video frames as input, a 3D convolutional neural network is used to extract inter-frame motion features, calculate optical flow information between adjacent frames, and generate a temporal attention weight map. The weight value is positively correlated with the pixel's motion amplitude;

[0082] 3) Spatiotemporal Joint Attention Fusion: A learnable weighted fusion strategy is adopted to fuse the spatial attention weight map and the temporal attention weight map to generate the final spatiotemporal joint attention weight map. The fusion formula is:

[0083]

[0084] in, and For learnable fusion weight coefficients, satisfying Automatic optimization through training;

[0085] 4) Attention weight smoothing: A 3×3 Gaussian kernel is used to smooth the spatiotemporal joint attention weight map to eliminate weight abrupt changes and ensure a natural transition of the enhancement effect.

[0086] Hierarchical Adaptive Enhancement Module

[0087] The video frame is divided into three regions: the core critical region ( Secondary key areas ( ) and background area ( Differentiated enhancement strategies are adopted for different regions:

[0088] 1) Enhancement of core critical areas: An adaptive histogram equalization algorithm based on guided filtering is used to improve contrast, combined with an unsharpened mask algorithm to enhance details and textures, while a slight brightness increase is also performed. The enhancement intensity is positively correlated with the attention weight.

[0089] 2) Enhancement of secondary critical areas: A linear contrast stretching algorithm is used to improve contrast, with appropriate brightness adjustment and noise suppression. The enhancement intensity is lower than that of the core critical areas.

[0090] 3) Background area enhancement: A gamma correction algorithm is used to make slight brightness adjustments to maintain the natural state of the background and avoid noise amplification and color distortion caused by over-enhancement;

[0091] 4) Pixel-level enhancement mapping: For each pixel, the final enhancement parameters are calculated based on its corresponding spatiotemporal joint attention weight value to achieve pixel-level adaptive enhancement. The brightness enhancement formula is:

[0092]

[0093] in, This is the original pixel brightness value. To enhance the brightness value of the subsequent pixels, This is the brightness enhancement factor, with a value range of [0.2, 0.5].

[0094] Video post-processing module

[0095] 1) Inter-frame consistency constraint: Calculate the difference in the mean brightness of adjacent frames after enhancement. If the difference exceeds a preset threshold (e.g., 5 gray levels), then the brightness of the next frame is linearly adjusted to eliminate the brightness jump between frames.

[0096] 2) Color correction: Slightly adjust the chroma components U and V of the enhanced video frame to maintain the naturalness and saturation of the colors and avoid color shift after enhancement;

[0097] 3) Format conversion and encoding: Convert the enhanced YUV format video frames to RGB format, compress and encode them using the H.264 or H.265 encoding standard, and output the enhanced short video file.

[0098] The present invention will be further described in detail below with reference to specific embodiments. This embodiment takes a 1080p, 30fps short video from a certain short video platform as the application object, and elaborates on the specific implementation process of the present invention.

[0099] I. System Deployment and Hardware Configuration

[0100] Preprocessing module: Deployed on the mobile CPU, using the ARM Cortex-A78 architecture, used for video decoding and frame preprocessing;

[0101] Key area detection module: Deployed on the mobile NPU, it adopts a lightweight detection model to achieve rapid detection of key areas in multiple dimensions;

[0102] Spatiotemporal attention modeling module: Deployed on the mobile NPU, it uses a quantized lightweight convolutional network to achieve rapid generation of attention weights;

[0103] Enhancement and post-processing module: Deployed on the mobile CPU, it uses SIMD instructions for acceleration to achieve layered enhancement and inter-frame consistency processing.

[0104] II. System Operation Flow

[0105] 1) Video preprocessing: Decode the input 1080p, 30fps short video, extract the original video frame sequence in YUV420 format, and normalize the luminance component to the [0,1] interval after Gaussian denoising; 2) Key region detection: Sequentially detect the face, text, product and moving object regions in the video frame, generate the initial key region mask, and the detection time is only 8ms / frame;

[0106] 3) Spatiotemporal attention modeling: A spatiotemporal joint attention weight map is generated through a spatiotemporal dual-branch network, which integrates the spatial key region distribution and temporal motion information, taking 5ms / frame;

[0107] 4) Layered adaptive enhancement: Differentiated enhancement is applied to different regions based on attention weights, with key enhancements to details in the core region and slight adjustments to the background region, taking 4ms / frame;

[0108] 5) Post-processing and encoding: Perform inter-frame consistency constraints and color correction to eliminate flickering, and finally use H.265 encoding to output the enhanced short video.

[0109] III. Implementation Results

[0110] This system was tested on a mobile device equipped with a Snapdragon 8 Gen2 processor. Enhancement processing was performed on 100 short videos in different scenarios (including indoor, outdoor, backlight, and low-light conditions). The test results are as follows:

[0111] Key area detection accuracy: 92.5%, capable of accurately identifying various key areas such as faces, text, products, and moving objects;

[0112] Subjective Quality of Life (MOS) score improved from 3.2 to 4.0, an increase of 0.8 points.

[0113] Key area information recognition accuracy improved by 42%, and text recognition accuracy increased from 75% to 96%.

[0114] Single frame processing speed: 32fps, meeting real-time processing requirements;

[0115] Inter-frame flicker rate: reduced from 15% to 2%, significantly improving video smoothness.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A short video key region enhancement system based on visual attention mechanism, characterized in that, include: The video preprocessing module is used to decode, extract frames and convert formats of the input short video to generate a continuous sequence of raw video frames, and perform noise reduction and normalization preprocessing. The key region detection module integrates face detection, text recognition, object detection and motion estimation technologies to perform multi-dimensional analysis on each original video frame, identify all key regions and generate an initial region mask. The spatiotemporal attention modeling module is used to construct a spatiotemporal dual-branch visual attention network, which combines the key region distribution in the spatial dimension with the inter-frame motion information in the temporal dimension to generate a pixel-level spatiotemporal joint attention weight map. The hierarchical adaptive enhancement module is used to divide video frames into core key regions, secondary key regions and background regions based on the spatiotemporal joint attention weight map, and to apply different enhancement algorithms and parameters to different regions to achieve hierarchical adaptive enhancement. The video post-processing module is used to perform inter-frame consistency constraints, color correction, and format conversion on the enhanced video frame sequence, and finally encode and output the enhanced short video.

2. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The video preprocessing module is specifically used to decode the input short video, extract the original video frame sequence in YUV format, perform Gaussian filtering to denoise the luminance component, and normalize the pixel values ​​to the [0,1] range.

3. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The key region detection module identifies key region types including face regions, text regions, product regions, and moving object regions. The logic for generating the initial region mask is as follows: corresponding region masks are generated by using face detection algorithm, text recognition algorithm, object detection algorithm, and motion estimation algorithm respectively, and the initial key region mask is obtained by performing a logical OR operation on all region masks.

4. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The spatiotemporal bi-branch visual attention network includes a spatial attention branch and a temporal attention branch. The logic for generating the spatiotemporal joint attention weight map is as follows: The spatial attention branch takes the original video frame of the current frame and the initial key region mask as input, and extracts multi-scale spatial features through a convolutional neural network to generate a spatial attention weight map. The temporal attention branch takes the current frame and the previous N original video frames as input, extracts inter-frame motion features through a 3D convolutional neural network, and generates a temporal attention weight map. A learnable weighted fusion strategy is adopted to fuse the spatial attention weight map and the temporal attention weight map to generate a spatiotemporal joint attention weight map, which is then Gaussian smoothed.

5. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The fusion formula for the spatiotemporal joint attention weight graph is: in, This is a spatial attention weight map. For time-based attention weights, and For learnable fusion weight coefficients, satisfying .

6. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The hierarchical adaptive enhancement module divides video frames into core key regions ( Secondary key areas ( ) and background area ( For the core critical areas, adaptive histogram equalization and unsharpened mask enhancement based on guided filtering are used; for the secondary critical areas, linear contrast stretching enhancement is used; and for the background areas, gamma correction enhancement is used.

7. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The formula for pixel-level brightness enhancement is: in, This is the original pixel brightness value. To enhance the brightness value of the subsequent pixels, This is the brightness enhancement factor. The spatiotemporal joint attention weight value is the corresponding pixel.

8. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The inter-frame consistency constraint logic of the video post-processing module is as follows: calculate the average difference in brightness between adjacent frames after enhancement; if the difference exceeds a preset threshold, then linearly adjust the brightness of the next frame to eliminate inter-frame brightness jumps and flickering.

9. The short video key region enhancement system based on visual attention mechanism according to claim 1, characterized in that, The key region detection module uses MTCNN or RetinaFace for face detection, CRNN or PaddleOCR for text recognition, YOLOv5 or YOLOv8 for object detection, and optical flow or frame difference for motion estimation.

10. An application method for a short video key region enhancement system based on visual attention mechanism as described in any one of claims 1-9, characterized in that, Includes the following steps: S1, Video Preprocessing: Decode, extract, denoise, and normalize the input short video to generate the original video frame sequence; S2, Key Region Detection: Integrates multi-dimensional detection technology to identify all key regions in the video frame and generate an initial region mask; S3, Spatiotemporal attention modeling: Pixel-level spatiotemporal joint attention weight map is generated through a spatiotemporal bi-branch visual attention network; S4, Layered Adaptive Enhancement: Differentiated enhancement is applied to different regions based on the attention weight map to achieve precise enhancement of key regions; S5, video post-processing: performs inter-frame consistency constraints, color correction and encoding, and outputs enhanced short videos.

Images