QP-adaptive compressed video super-resolution with coding priors

Abstract

Video super-resolution algorithms have found widespread applications as post-processing techniques in down-sampling based coding methods. With the advancements in deep learning techniques, video super-resolution has achieved remarkable success. However, applying existing video super-resolution methods to compressed videos requires training specific models for various quantization parameters (QPs), significantly increasing the resource consumption for model training and compromising their practical utility. To address this issue, we propose a QP-adaptive network for compressed video super-resolution based on coding priors (QPAN). Firstly, we design a QP modulation module (QPMM), which can utilize the frame-wise QP to recalibrate feature maps. Then, on the basis of QPMM, an adaptive multi-scale prior fusion module (Ada-MSPFM) and an adaptive enhancement modulation module (Ada-EMM) are constructed. The former effectively integrates multi-scale features from spatial coding priors in the bitstream and multi-scale features from the decoded video frames. And the latter improves the expressive ability of the network by leveraging QP modulation and reinforcing feature flow adaptively. Extensive experiments demonstrate the highly flexible and adaptive of our proposed method, which exhibits superior reconstruction performance compared to state-of-the-art video super-resolution algorithms.