Knowledge Distillation-Based Lightweight Image Compression Method

With the rapid development of mobile internet, the increasing demands for image transmission and storage require advanced compression technologies to accelerate transmission speed and reduce storage costs. This paper proposes a feature distillation learning framework based on Correlation Congruence and Similarity-Preserving Knowledge Distillation (SPKD) method, integrating joint autoregressive hierarchical priors model for lightweight image compression. To address the bias caused by feature inconsistency in traditional feature distillation approaches, we replace the tensor in feature learning with its orthogonal projection. Experimental results demonstrate that our improved distillation method achieves PSNR (28.31) and MS-SSIM (0.93), outperforming direct training methods (26.90 PSNR and 0.91 MS-SSIM) and showing 2.97 and 0.04 improvements over static knowledge distillation methods in these two metrics, respectively. The study reveals that under equivalent PSNR conditions, our student model reduces BPP (bits per pixel) by 18% compared to JPEG2000. These results validate the feasibility of our knowledge distillation-based compression framework.