Lightweight Video Frame Interpolation Based on Bidirectional Attention

• 1、Information Science and Engineering School, Hunan University, Changsha 410082

Abstract：\justifying Video frame interpolation is an important topic in the field of computer vision. It aims to improve the frame rate of a video by synthesizing non-existent intermediate frames between two consecutive frames, thus enhancing the continuity and smoothness of the video. Existing algorithms have difficulty in synthesizing high quality interpolation frames under scenarios such as occlusion and motion blurs. Meanwhile, most interpolation methods are unable to balance the performance and complexity of the model.The number of parameter and computation is enormous in exchange for higher objective evaluation metrics. To this end, this paper proposes a lightweight network combined with bidirectional attention module that effectively combining channel attention and spatial attention, making the model better capture variable region between the input frames and enhancing features representation related to motion. Futhernore, a lightweight Ghost module is introduced to the model which can synthesize features at a lower cost, decreasing the comsuption of memory and computation resource. Experimental results on Middlebury dataset show that the peak-signal to noise ratio is inproved by 0.06dB, and the number of parameter is only about one tenth of the original, verifing the effectiveness of our method.

Keywords： Video Frame Interpolation Deep Learning Attention Mechanism Model Compression