Learned Image Compression with Lightweight Attention

Recently many learned image compression methods have surpassed traditional techniques in term of rate-distortion performance. However, these methods have yet to effectively eliminate local redundancy, and their complex structures and high computational costs also prevent their wide applications to power-sensitive, resource-limited devices, such as mobile phones. To resolve those issues, we propose a learned image compression approach that leverages a lightweight attention mechanism. Specifically, the proposed Local-Nonlocal Attention Module (LNAM) parallels the processing of local and non-local features, thereby eliminating redundancy more efficiently. Moreover, we further introduce a lightweight channel-wise entropy model to reduce the number of bits required for encoding. Experimental results demonstrate that our method achieves a favorable tradeoff among computational cost, latency, and rate-distortion performance across three public datasets, including Kodak, CLIC, Tecnick.