Adaptive cross-channel transformation based on self-modulation for learned image compression

Recently learned image compression has achieved excellent rate–distortion performance, and nonlinear transformation becomes a critical component for performance improvement. While Generalized Divisible Normalization (GDN) is a widely used method that exploits channel correlation for effective nonlinear representation, its utilization of cross-channel relationship for each element of features remains limited. In this paper, we propose a novel cross-channel transformation based on self-modulation, named SMCCT. The SMCCT takes the intermediate feature maps as input to capture cross-channel correlation and generate affine transformation parameters for element-wise feature modulation. The proposed transformation enables adaptive weighting and fine-grained control over the features, which helps to learn expressive features and further reduce redundancies. The SMCCT can be flexibly employed into learned image compression models. Experimental results demonstrate that the proposed method can achieve superior rate–distortion performance with the existing learned image compression methods and outperform traditional codecs under the quality metric such as PSNR and MS-SSIM. Specifically, when using the PSNR metric, our proposed method outperforms latest codec VTM-12.1 by 5.47%, 10.25% in BD-rate on Kodak and Tecnick datasets. When using the MS-SSIM metric, it outperforms latest codec VTM-12.1 by 50.97%, 49.81% in BD-rate on Kodak and Tecnick datasets.