Dual-stage feedback network for lightweight color image compression artifact reduction.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Lossy image compression introduces artifacts, degrading visual quality.
• Deep convolutional neural networks show promise for artifact reduction but often neglect chroma channels and suffer from high complexity.
• Existing methods struggle with practical deployment due to computational demands.

Purpose of the Study:

• To develop a lightweight network for effective color image compression artifact reduction.
• To address the limitations of existing methods by considering both luma and chroma components.
• To propose a novel network architecture with reduced model complexity for practical applications.

Main Methods:

• A Dual-Stage Feedback Network (DSFN) is proposed for lightweight color image compression artifact reduction.
• A curriculum learning strategy guides the DSFN in a luma-to-RGB manner, focusing on luma reconstruction first.
• An enhanced feedback block with adaptive iterative self-refinement and separable convolutions facilitates efficient feature extraction.

Main Results:

• The DSFN demonstrates significant advantages over state-of-the-art methods in quantitative metrics and visual quality.
• The proposed method achieves notable compression artifact reduction, including for chroma components.
• The DSFN exhibits lower model complexity, enhancing its practicality.

Conclusions:

• The DSFN offers an effective and efficient solution for color image compression artifact reduction.
• The luma-to-RGB approach and novel feedback mechanism contribute to improved restoration quality.
• The lightweight design makes the DSFN suitable for real-world deployment.