An Enhanced Channel-Wise Autoregressive Learned Image Compression Model With NLSR, Transformer, and PPCA

Abstract

This paper presents an enhanced deep learning–based image compression framework aimed at improving rate–distortion performance and reconstruction quality. The proposed method is built upon a hyperprior-based architecture and integrates several key components, including a Non-Local Super-Resolution (NLSR) module, a Transformer (TF) block, and a Probabilistic Principal Component Analysis (PPCA) block. These components are designed to improve latent representation, refine feature extraction, and recover high-frequency image details. In addition, a denoising module is applied as a post-processing step to further suppress compression artifacts and enhance visual quality without increasing the bitrate. The NLSR module enhances spatial detail recovery during reconstruction, while the TF block improves contextual modeling and enables more accurate estimation of entropy parameters. The PPCA block further refines latent features through probabilistic dimensionality reduction, leading to improved reconstruction fidelity. The denoising stage effectively reduces residual noise and enhances texture clarity. Experimental results demonstrate that the proposed framework consistently outperforms existing methods in both objective metrics and perceptual quality. The overall design achieves a better balance between compression efficiency and visual fidelity, particularly in images with complex textures, making it a robust solution for high-quality image compression applications.