Optimizing Image Compression: Perspectives on SVD-Based Restoration

Recently, neural image codecs have demonstrated impressive performance in image compression. Current approaches primarily focus on designing sophisticated context mechanisms or network architectures for eliminating redundancies within image data, which leads to more and more computational resource requirements and increases the encoding and decoding time. In this paper, we analyze the degradation of compressed images from the perspective of Singular Value Decomposition (SVD) and propose a novel rate-distortion (RD) optimization framework based on compressed image restoration. Specifically, our method introduces an SVD-based basis reconstruction error into the conventional R-D loss function, enabling enhanced detail restoration capabilities. The proposed optimization framework can be seamlessly applied to various codecs optimized by the R-D loss without introducing additional learnable parameters or inference overhead. Extensive experiments on public datasets show that our method achieves performance gains across several neural image codecs compared to the baseline, validating the effectiveness of the proposed optimization framework.