An Unsupervised Domain Adaptation Method for Compressed Video Quality Enhancement
Zeyang Wang · 2022 19th International Computer Conference on Wavelet Active Media Technology and Information Processing (ICCWAMTIP) · 2022
In recent years, the field of deep learning based compressed video quality enhancement (CVQE) has developed rapidly. However, due to the domain shift, the model struggles to perform effectively on those videos which has inconsistent distribution with training sets. Recently, a domain adaptation method called mean teacher has been proposed in the field of object detection. The mean teacher technique has already had a lot of success with cross-domain recognition. In this paper, we introduce the mean teacher method to CVQE, interfere with the output of teacher and student model by adding noise, and use consistency loss to guide the training of student model. This is the first work of source-free domain adaptation on CVQE. Our experiment demonstrates the practicality of the unsupervised approach on CVQE.