Multi-level feature interaction image super-resolution network based on convolutional nonlinear spiking neural model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Single Image Super-Resolution (SISR) aims to enhance low-resolution images to high-resolution.
• Existing Convolutional Neural Network (CNN) methods struggle to integrate global and local image features effectively.
• Current models often overlook correlations between hierarchical feature information.

Purpose of the Study:

• To propose a novel multi-level feature interactive image super-resolution network.
• To address limitations in combining global/local information and inter-hierarchical feature correlations in SISR.
• To develop a network architecture inspired by nonlinear spiking neural P systems.

Main Methods:

• Developed a network with shallow feature processing, deep feature extraction/fusion, and reconstruction modules.
• Incorporated omni-domain self-attention blocks for global information extraction.
• Designed a novel convolutional unit and a multi-level feature fusion module for local information extraction and adaptive feature fusion.

Main Results:

• The proposed model was compared against 16 state-of-the-art and baseline models on five benchmark datasets.
• Experimental results demonstrated superior reconstruction performance.
• The model achieved a favorable balance between performance and computational complexity (model parameters).

Conclusions:

• The multi-level feature interactive network effectively enhances single image super-resolution.
• The integration of self-attention and novel convolutional units improves feature extraction and fusion.
• The proposed approach offers a promising solution for high-quality image super-resolution with efficient parameter usage.