Towards real-time practical image compression with lightweight attention

Abstract

The rate–distortion (RD) performance of learning-based image compression (LIC) has already surpassed that of traditional Versatile Video Coding (VVC) intra-coding. However, the gain of the compression efficiency is at the cost of high computational complexity, which is prohibitively expensive and becomes a bottleneck for practical applications. To this end, we proposed an efficient LIC method with lightweight designs for real-time practical applications, which achieves a better trade-off between compression performance and computational complexity. Specifically, residual connected lightweight attention units (RLAUs) are stacked for feature extraction, which achieves effective global spatial information embedding while keeping relatively low complexity. Meanwhile, a trainable channel-gained adaptive module (CGAM) is introduced into the nonlinear transform network and multi-stage context model. This module re-distributes the importance of different channels, further improving compression efficiency. The experimental results demonstrate that the proposed method achieves better compression efficiency than VVC intra-coding. Furthermore, compared to other state-of-the-art LIC approaches, the proposed method significantly reduces the coding time while maintaining comparable coding efficiency. The source codes are available at https://github.com/llsurreal919/LightweightLIC.