Content-Adaptive Acceleration for Block-based Learned Image Compression

Learned Image Compression has made great progress recently. However, most existing works process entire images in inference, which may cause out-of-memory errors when with high-resolution images or complex models. To limit peak memory usage, a viable way is to divide an image into blocks and process them separately. We found that blocks with different contents can be represented with varying levels of compactness, e.g., flatter blocks can be more compactly represented to accelerate entropy coding. Thus we propose ClassLIC, which consists of three modules, to achieve content-adaptive acceleration. In detail, Class-Module classifies input blocks into different categories and LIC-Module uses different models to process different categories. An optional Deblock-Module helps mitigate blocking artifacts that may occur in the combined image at low bitrates. Experiments show that ClassLIC reduces coding time to 35.3% of the block-based method at most while maintaining compression efficiency.