PSAR-SR: Patches separation and artifacts removal for improving super-resolution networks

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Existing super-resolution (SR) methods like ClassSR decompose images into patches for efficient processing.
• However, ClassSR faces challenges including increased training difficulty and artifacts due to overlapping patches.
• These limitations hinder the effectiveness of large image SR.

Purpose of the Study:

• To propose an end-to-end general framework, PSAR-SR, for efficient and artifact-free large image super-resolution.
• To address the challenges of varying patch recovery difficulty and artifacts in image decomposition.
• To reduce computational cost while enhancing SR performance.

Main Methods:

• Developed an Image Information Complexity Module (IICM) to assess patch recovery difficulty.
• Introduced a Patches Classification and Separation Module (PCSM) for dynamic SR path selection.
• Implemented a Multi-Attention Artifacts Removal Module (MARM) for artifact reduction and computational efficiency.
• Proposed Threshold Penalty Loss (TP-Loss) and Artifacts Removal Loss (AR-Loss) to optimize SR path selection and reconstruction quality.

Main Results:

• PSAR-SR effectively eliminates artifacts in overlapping-free decomposition.
• Achieved superior performance compared to leading SR methods (FSRCNN, CARN, SRResNet, RCAN, CAMixerSR).
• Demonstrated significant computational cost savings, reducing FLOPs by 53%-65%.

Conclusions:

• PSAR-SR offers an effective solution for large image super-resolution by managing patch complexity and artifacts.
• The framework achieves a strong balance between performance enhancement and computational efficiency.
• The proposed methods and loss functions contribute to improved SR quality and reduced artifacts.