The Design And Implementation Of UGC Short Video Quality Assessment System Based On Deep Learning

At present,short video products have developed into one of the most popular Internet products.In major short video platforms,a large number of UGC(User Generated Content)short videos are produced every day.Relying entirely on manual review of video quality is not only costly,but also inefficient,which seriously hinders the distribution and dissemination of high-quality videos.To solve this problem,this thesis proposes a UGC video quality assessment algorithm based on deep learning,and builds a UGC short video quality assessment system based on the algorithm.The algorithm proposed in this thesis uses CNN and LSTM network to capture spatial distortion information and temporal distortion information of UGC videos,respectively.Unlike previous work,in order to extract the deep features related to spatial distortion in UGC videos more accurately,we train a Siamese network that can compare the relative image quality based on a large number of images with different types and different degrees of distortion,and use one of its sub-networks as a pre-training network to fine-tune it on the UGC video quality dataset.After extracting the deep features of each UGC video,we input the sequences of frame-level deep features into the LSTM network and train it so that it can predict the final quality of UGC videos.The system built in this thesis creates a video quality review mode with machine review as the main and manual review as the supplement.Among them,machine review is mainly used to filter low-quality videos quickly,manual review is mainly used to evaluate the results of machine review.In the machine review mode,super reviewers can train the machine review models,and use the trained models to predict the quality scores of the UGC short videos.In the manual review mode,super reviewers can distribute manual review tasks to ordinary reviewers,and ordinary reviewers can rate the quality of the videos in the task.The system is developed based on the Vue and Flask,and is deployed through Docker.On the Konvid-1k dataset,the algorithm achieves PLCC and SROCC metrics of0.788 and 0.789 respectively,which is a significant improvement over other algorithms.In addition,the results of the ablation and cross dataset experiments demonstrate the necessity of each step in the algorithm and the generalization ability of the algorithm respectively.The test results show that the system meets all functional and non-functional requirements.