Research And Application Of Compressed Video Quality Enhancement Based On Deep Learning

In recent years,global video data has grown rapidly.However,due to limitations in network bandwidth and storage space,videos are often saved in a lossy compressed format.Video encoding and decoding algorithms greatly reduce video bit rates but also introduce irreversible quality losses.Thus,improving compressed video quality has become an urgent issue.The development of deep learning technology has led many scholars to use this technique to enhance compressed video quality.However,current technology still faces some challenges,such as insufficient use of spatial information and difficulty in industrial applications.This thesis proposes two algorithms based on existing deep learning technology to address these problems and applies them to a surveillance system.The main contributions of this thesis can be summarized as follows:First,the thesis proposes the concept of balancing spatial and temporal information to address the unified modeling challenge of peak quality frames and non-peak quality frames.Based on this,an off-the-shelf spatial-temporal information balance module is designed.The module refines temporally aligned features using spatial attention networks and balances the proportion of spatial and temporal information using channel attention mechanisms and re-alignment.This module is applicable to existing multi-frame quality enhancement architectures based on pre-alignment.The thesis combines this module with existing multi-frame video quality enhancement models and conducts experiments.The results show that the proposed method outperforms existing representative multi-frame quality enhancement methods.Second,the thesis proposes an unsupervised domain adaptation method for compressing video quality enhancement models.This method can update the model with only pre-trained models and unlabeled target domain videos,and improves the enhancement effect of the model in the target domain.In transfer experiments at the same QP and different QP levels,we demonstrate the feasibility of this method and draw three important conclusions:(1)manual classification cannot well represent the actual distribution differences between video data?(2)under the same QP,the higher the QP,the better the transfer effect?(3)under different QP levels,transferring to higher QP levels is more effective than transferring to lower QP levels.Third,a client/server-based intelligent video surveillance system is designed and implemented.The system includes video acquisition,encoding,transmission,decoding,and playback functions.The two algorithms proposed in this thesis are applied to the system,realizing real-time enhancement of surveillance videos and online learning updates of the model.In summary,this thesis successfully demonstrates the effectiveness of the spatialtemporal information balance module,reveals the problem of insufficient use of spatial information in existing multi-frame enhancement methods,and suggests that future research should consider how to better balance spatial and temporal information.Furthermore,the thesis proposes for the first time an unsupervised domain adaptation method,which is successfully applied to the quality enhancement field,demonstrating its great potential.In addition,the thesis designs and implements an intelligent video surveillance system,further verifying the feasibility and practicality of the proposed methods.