Temporal Adaptive Learned Surveillance Video Compression

Abstract

As the amount of surveillance video data increases at an exponential rate, the need for efficient video compression algorithms becomes increasingly urgent. The inter-frame compression schemes of existing surveillance video compression methods predict the current frame through the previous frame, causing the error to gradually increase because the quality of the reference frame decreases progressively. In this paper, we propose a Temporal Adaptive enhancement method for Learned surveillance video Compression (TALC). The proposed TALC has two modules: Forward Temporal Adaptive (FTA) module and Backward Temporal Adaptive (BTA) module which are put before and after motion and residual bits transmission modules respectively. These two modules have the same network structure which consists of a Temporal Adaptive Selection (TAS) block and a Feature Enhancement (FE) block. TAS block can analyze the extent which errors accumulate in optical flow and residuals, then select the corresponding enhancement sub-block; while FE block consists of several enhancement sub-blocks according to different levels of error accumulation. The proposed TALC has strong versatility and low coupling, which can be applied in almost all learned video compression frameworks as a plugin. Experimental results show that the proposed TALC method can significantly improve the coding performance of learned surveillance video compression networks without changing the original basic structure.