BSTN: An Effective Framework for Compressed Video Quality Enhancement

In this paper, we propose a boundary-guided spatial-temporal network (BSTN) to enhance the quality of HEVC compressed videos. We first adopt the advanced optical flow algorithm to estimate the motion between the current frame and its frames, these adjacent frames are pre-warped by the estimated motion flow. Then, a bi-directional residual convolutional LSTM (BRCLSTM) is designed to implicitly discover frame variations over time between the compensated frames and current frame. In addition, we generate a guided map by utilizing the partition information of transform units (TUs) to guide our network to concentrate more on the boundaries of TUs, and the guided map is fused into our network by a boundary attention fusion module to enhance the reconstructed quality of current frame. Finally, some stacked convolutional layers with ReLU are used to reconstruct the desired frame. We evaluate our method under different configurations. Experimental results show that our method can produce superior results than the state-of-the-arts, both objectively and subjectively.