RefQSR: Reference-Based Quantization for Image Super-Resolution Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep learning models for Single Image Super-Resolution (SISR) achieve high performance but are computationally expensive.
• Network quantization offers a solution for efficient SISR, but existing methods do not leverage image self-similarity.
• Resource-constrained environments necessitate computationally efficient SISR solutions.

Purpose of the Study:

• To develop a novel network quantization method for SISR that exploits image self-similarity.
• To enhance the computational efficiency of SISR models without sacrificing performance.
• To introduce a new approach for applying quantization in image super-resolution.

Main Methods:

• Proposed Reference-Based Quantization for Image Super-Resolution (RefQSR) method.
• Developed dedicated patch clustering and reference-based quantization modules.
• Integrated RefQSR into existing SISR network quantization frameworks.

Main Results:

• RefQSR effectively utilizes image self-similarity for quantization.
• The method demonstrates significant improvements in computational efficiency for SISR.
• Experimental results validate RefQSR's effectiveness across various SISR networks and quantization techniques.

Conclusions:

• RefQSR offers an effective strategy for efficient SISR through reference-based quantization.
• The proposed method successfully exploits image self-similarity, a novel direction for SISR quantization.
• RefQSR provides a viable solution for deploying SISR in resource-limited settings.