HIPA: Hierarchical Patch Transformer for Single Image Super Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Transformer architectures show promise in Single Image Super-Resolution (SISR).
• Existing methods have limitations in patch size adaptability and token dependency handling.
• Fixed patch sizes and uniform position encodings hinder optimal performance on diverse image regions.

Purpose of the Study:

• To address limitations in current Transformer-based SISR methods.
• To introduce a novel Transformer architecture, HIPA, for enhanced super-resolution.
• To improve the adaptive restoration of image textures and details.

Main Methods:

• Developed a Hierarchical Patch Partition (HIPA) Transformer architecture.
• Implemented a cascaded model processing images in multiple stages with progressively merging tokens.
• Introduced an attention-based position encoding scheme and a multi-receptive field attention module.

Main Results:

• HIPA demonstrates superior performance in quantitative and qualitative evaluations on public datasets.
• The hierarchical patch mechanism adaptively learns features for different image regions.
• Attention-based position encoding effectively assigns weights to tokens, improving dependency handling.

Conclusions:

• HIPA offers a significant advancement in Transformer-based Single Image Super-Resolution.
• The proposed hierarchical and adaptive approach enhances the restoration of fine details and rich textures.
• HIPA provides a new state-of-the-art for SISR tasks.