Multimedia Systems Компьютерные сети и коммуникации

Zero-shot JPEG image deblocking: leveraging non-local self-similarity for enhanced restoration

Gu, Feijun, Yu, Zhan, Bai, Xiuyue, Xie, Mingzheng, Wang, Zijing, Zhou, Xin
Springer Nature · 2026-11-13 · DOI 10.1007/s00530-026-02777-x

Аннотация

JPEG compression produces visually distracting blocking artifacts due to the independent coding of $$8\times 8$$ blocks. This paper presents a dataset-free, single-image internal-learning method for JPEG artifact removal. For each input image, a compact residual convolutional network is optimized at test time using self-supervision constructed from non-local patch correspondences. Specifically, patches with different JPEG-grid phases are retrieved and intensity-aligned to form a confidence-weighted consensus target. The resulting supervision is imposed selectively on pixels near coding-block boundaries, thereby reducing blocking artifacts while limiting undesirable modifications in content regions. In addition, a content-preserving loss and an input-only weak-compression regularizer are introduced to mitigate over-smoothing and retain fine image details. Experiments on Set5, Set14, LIVE1, and Classic5 are conducted under a unified JPEG compression protocol. The proposed method achieves image-count-weighted mean PSNR values of 27.66, 30.37, and 32.33 dB at quality factors of 10, 20, and 40, respectively. With an identical budget of 800 test-time optimization iterations, it attains the best PSNR among the compared single-image baselines across all 12 dataset–quality-factor settings, and the best SSIM in 11 settings. The model contains only 19,105 trainable parameters and requires approximately 4 s per image on an NVIDIA RTX 4070 GPU.

Читать у издателя →