Video Super-Resolution With Multi-Level Attention

Abstract

A video super-resolution reconstruction algorithm with multi-level attention is proposed to fully utilize the abundant inter-frame redundancy information and balance the trade-off between reconstruction performance and speed. The algorithm achieves adaptive residual space fusion and multi-level inter-frame information fusion through a recursive adaptive aggregation network and adaptive multi-level attention modules. Specifically, each adaptive multi-level attention module is used for multi-level inter-frame information fusion. Then, multiple cascaded adaptive multi-level attention modules with shared weights are used for adaptive residual space fusion. Finally, the feature is refined and enlarged through a reconstruction network to obtain the final high-resolution video frame. This algorithm can better restore high-frequency details such as textures and edges, while enhancing the long-term temporal dependency modeling capability and achieving a balance between reconstruction performance and speed. Experimental results show that the proposed algorithm can effectively improve the video super-resolution reconstruction performance on standard datasets.