Efficient progressive training with granularity cross for image super-resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deeper models yield better super-resolution results but are difficult to train.
• Large deep models pose significant training challenges in image super-resolution.

Purpose of the Study:

• To propose an efficient progressive training framework with granularity cross (EPTGC) for image super-resolution.
• To address the difficulties in training large deep image super-resolution models.
• To enhance the recovery of image edge information.

Main Methods:

• Propose EPTGC, a plug-and-play framework that splits models and uses multi-granularity images.
• Apply EPTGC to 8 diverse image super-resolution models (CNN and transformer-based).
• Evaluate EPTGC on 4 benchmark datasets.

Main Results:

• EPTGC reduces training difficulty for large deep models.
• The framework enhances the model's ability to learn granular features and recover image edges.
• Significant improvements observed, with a maximum PSNR gain of 0.44 dB on benchmark datasets.

Conclusions:

• EPTGC is an effective approach to improve image super-resolution model training and performance.
• The method enhances edge information recovery, a critical aspect of super-resolution.
• EPTGC offers a versatile solution applicable to various existing super-resolution architectures.