Exploring High Dimensional Feature Space With Channel-Spatial Nonlinear Transforms for Learned Image Compression

Nonlinear transforms have significantly advanced learned image compression (LIC), particularly using residual blocks. This transform enhances the nonlinear expression ability and obtain compact feature representation by enlarging the receptive field, which indicates how the convolution process extracts features in a high dimensional feature space. However, its functionality is restricted to the spatial dimension and network depth, limiting further improvements in network performance due to insufficient information interaction and representation. Crucially, the potential of high dimensional feature space in the channel dimension and the exploration of network width/resolution remain largely untapped. In this paper, we consider nonlinear transforms from the perspective of feature space, defining high-dimensional feature spaces in different dimensions and investigating the specific effects. Firstly, we introduce the dimension increasing and decreasing transforms in both channel and spatial dimensions to obtain high dimensional feature space and achieve better feature extraction. Secondly, we design a channel-spatial fusion residual transform (CSR), which incorporates multi-dimensional transforms for a more effective representation. Furthermore, we simplify the proposed fusion transform to obtain a slim architecture (CSR-sm), balancing network complexity and compression performance. Finally, we build the overall network with stacked CSR transforms to achieve better compression and reconstruction. Experimental results demonstrate that the proposed method can achieve superior rate-distortion performance compared to the existing LIC methods and traditional codecs. Specifically, our proposed method achieves 9.38% BD-rate reduction over VVC on Kodak dataset.