UGC Videos Enhancement And Super-resolution And Videos Classification

With the development of multimedia and communication technology,the amount of video data has been rapidly emerging.More and more people choose to shoot video and upload it to the network to share their life.We called these videos as user generated content(UGC)video and these videos are tend to be low-quality.Video enhancement and super-resolution(VESR)technology is drawing more and more attention in com-puter vision from both the research and industrial community.However,in existing VESR datasets,low-quality videos were generated via down-sampling from its high-quality counterpart,which cannot cover the realistic degradations such as random noise,compression artifact,and transmission distortion,and thus the VESR algorithms devel-oped based on these datasets fail in real-world user generated content video.To facilitate the research of VESR towards real applications including realistic user generated con-tent videos,we build a realistic and large-scale video enhancement and super-resolution dataset in the wild(called VESR-Wild),which 1)is collected from an industrial video service and covers realistic degradations;2)contains 10,000 video clips and totally 1 million video frames covering diverse video categories,and to the best of our knowl-edge,is the most diverse and largest dataset for VESR.We test and compare popular VESR methods on this VESR-Wild dataset and further discuss future research directions for VESR.We hope that our dataset can help people to gain deeper understandings of existing VESR algorithms and enable the development of new algorithms for real-world UGC video applications.Then,in this paper,we introduce VESR-Net,a method for video enhancement and super-resolution(VESR).We design a separate non-local(Separate NL)module to explore the relations among video frames and fuse video frames efficiently,and a chan-nel attention residual block(CARB)to capture the relations among feature maps for video frame reconstruction in VESR-Net.We conduct experiments to analyze the effec-tiveness of these designs in VESR-Net,which demonstrates the advantages of VESR-Net over previous state-of-the-art VESR methods.It is worth to mention that among more than thousands of participants for Youku video enhancement and super-resolution(Youku-VESR)challenge,our proposed VESR-Net beat other competitive methods and ranked the first place.In real-world applications,VESR can be used with video classification.Consider-ing the affact of low-quality UGC videos on video classification,we conduct extensive experiments on the affect of current VESR algorithms on video classification.In addi-tion,it is still challenging in real applications due to motion blur,object occlusion and extreme illumination in UGC videos.We propose a multi-branch voting network and a UGC video classification dataset which facilitate the video classification in realistic scenarios.