LIC-PHM: Parameterized Hybrid Module for Learned Image Compression

Learned image compression has attracted great attention in image and video coding, due to its learnable capabilities and powerful expressiveness. However, the introduction of vision transformer seems to have limited the room for improvement in its coding performance. Inspired by feature representation, convolutional layers (named as CNN) can preserve the locality of image content and vision transformer layer can maintain the global attributes of image, but the latent features of them cannot describe the finer-detailed image content. In this paper, the parameterized hybrid module is proposed to obtain the finer representation of image contours and aggregate the multiple types of feature for efficient image compression. Firstly, we introduce the mixed CNN and Swin Transformer architecture as the base model, integrate the graph-structured Swin Transformer scheme and design a graph-aware edge convolution layer. Secondly, we combine the convolution branch, the two-stage graph-Transformer branch and graph-aware edge convolution branch in parameterized weighing manner to obtain the hierachical features of latent representation. Finally, a noval learned image compression with parameterized hybrid module is trained on 300K images of the OpenImage dataset, and the trained model is tested on Kodak dataset.