AnySR: Realizing Image Super-Resolution as Any-Scale, Any-Resource.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Single-Image Super-Resolution (SISR) methods often require significant computational resources, limiting their scalability and efficiency across different scales.
• Existing arbitrary-scale SISR solutions typically use the same computational cost regardless of the target scale, leading to inefficiencies.

Purpose of the Study:

• To introduce AnySR, a novel framework for rebuilding existing arbitrary-scale SISR methods into any-scale, any-resource implementations.
• To enhance the efficiency and scalability of SISR applications by optimizing resource utilization for different scales.

Main Methods:

• AnySR builds arbitrary-scale tasks as any-resource implementations, reducing resource demands for smaller scales without additional parameters.
• The framework enhances any-scale performance through a feature-interweaving approach, inserting scale pairs into features at regular intervals for correct processing.

Main Results:

• AnySR was demonstrated by rebuilding existing arbitrary-scale SISR methods and validated on five popular SISR test datasets.
• The results confirm that AnySR implements SISR tasks more efficiently and performs comparably to existing arbitrary-scale SISR methods.

Conclusions:

• AnySR achieves both any-scale and any-resource implementation for SISR tasks, a first in the field.
• This framework offers a more computing-efficient solution for SISR, making advanced super-resolution techniques more accessible across diverse hardware capabilities.