An efficient multi-scale learning method for image super-resolution networks.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image super-resolution (SR) faces challenges with one-to-many mappings from low-resolution (LR) to high-resolution (HR) spaces.
• Existing SR networks treat different scales as independent tasks, limiting feature reuse and increasing computation.
• Current arbitrary scale SR methods do not fully address these inefficiencies.

Purpose of the Study:

• To propose an efficient multi-scale learning method for image SR networks.
• To introduce a novel self-generating (SG) mechanism to enhance SR performance and reduce computational load.
• To develop a new SG upscale module to replace traditional upscaling methods.

Main Methods:

• The proposed SG-SR method utilizes learned features to generate upscale filters via a novel SG upscale module.
• The SG upscale module applies spatial weights to LR tensors, which are then processed to generate HR images.
• The approach enables efficient multi-scale feature learning and direct HR image reconstruction.

Main Results:

• Extensive experiments demonstrate superior performance of SG-SR compared to state-of-the-art (SOTA) methods on benchmark datasets.
• The SG upscale module effectively enhances the performance of existing SR networks.
• The proposed module achieves significantly lower computational cost than conventional upscale modules.

Conclusions:

• The SG-SR method offers an efficient and effective solution for multi-scale image super-resolution.
• The novel SG upscale module provides a computationally efficient way to improve SR performance.
• This work advances the field of image SR by enabling better feature utilization and faster high-resolution image recovery.