Jpeg stereo image lossy recompression with mutual information enhancement

• Published in: Multimedia systems Vol. 31; no. 5
• Main Authors: Zhou, Junwei, Zhang, Benyi, Wu, Shengping, Zhou, Lei, Yang, Yanchao, Xiang, Jianwen
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2025; Springer Nature B.V
• Subjects: Algorithms; Alignment; Compressing; Computer Communication Networks; Computer Graphics; Computer Science; Cryptology; Data integration; Data Storage Representation; Discrete cosine transform; Image compression; Modules; Multimedia Information Systems; Operating Systems; Regular Paper; Signal to noise ratio; Distributed coding; Mutual information enhancement; JPEG image lossy recompression
• ISSN: 0942-4962; 1432-1882
• DOI: 10.1007/s00530-025-01945-9

Despite JPEG remaining the most prevalent image compression algorithm, the majority of these algorithms primarily concentrate on uncompressed images, thereby overlooking the substantial quantity of existing compressed JPEG images. The recent strides in JPEG recompression strive to further minimize these JPEG file sizes. Unfortunately, the prevalent techniques, especially lossy recompression, frequently fail to fully utilize the correlation of coefficients, resulting in less-than-optimal compression. Moreover, they do not tap into the similarities among JPEG stereo images to boost compression. This paper proposes a novel approach, Distributed Coding for JPEG Lossy Recompression (DCLRC), to enhance the efficiency of JPEG recompression. DCLRC employs an encoding network to eliminate redundant data within the Discrete Cosine Transform (DCT) domain. This is followed by preprocessing via DCT coefficient reconstruction and alignment networks allocated for dimensionality reduction and channel alignment. More significantly, DCLRC capitalizes on the visually similar images in the DCT domain by incorporating a mutual information enhancement module. This pioneering module combines feature extraction, multi-head cross-attention, and information fusion. The experimental results show that DCLRC outperforms existing JPEG lossy recompression techniques, including ULIC, Balle2018, BPG, VCC-Intra, HEVC-Intra, JPEG2000, EILC2022, DSIN, and NDIC2022. DCLRC has achieved an average Peak Signal-to-Noise Ratio (PSNR) improvement of 2.0 dB as well as an average Multi-Scale Structural SIMilarity (MS-SSIM) improvement of 2.3 dB on the KITTI dataset at low bitrates. Such results unequivocally verify the effectiveness of DCLRC in augmenting JPEG recompression performance. Our code is available at https://github.com/wsp11/DCLRC.