A deep learning approach for quality enhancement of surveillance video

The growing number of surveillance cameras imposes great demand on high efficiency video coding. Although modern video coding standards have significantly improved the video coding efficiency, they are designed for general video rather than surveillance video. The special characteristics of surveillance video leave a large space for further performance improvement. In this paper, we leverage a deep learning approach to enhance the quality of compressed surveillance video. More specifically, we formulate the problem of frame enhancement as a regression problem and design a Residual Squeeze-and-Excitation Network (RSE-Net), to address it. RSE-Net extensively exploits the non-linear mapping between the reconstructed frame and the ground truth, with only a small number of parameters. Moreover, By improving You Only Look Once (YOLO) network, we successfully detect the grouped vehicles within a frame. A novel model training scheme is then developed through learning from the grouped vehicles. With the proposed scheme, we train a global model for both foreground and background of surveillance video. Experimental results show that our method achieves average 0.40 dB, 0.22 dB and 0.24 dB PSNR gains over H.265/HEVC anchor in AI, LDP and RA configurations, and produces visually pleasing results when applied to compressed surveillance video.