Quaternion attention multi-scale widening network for endoscopy image super-resolution.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep convolutional neural networks (CNNs) are used for endoscopic Super-Resolution (SR).
• Existing methods often have large parameters, limiting practical application.
• Current approaches treat image channels equally, ignoring inter-channel differences.

Purpose of the Study:

• To design a lightweight SR model named Quaternion Attention Multi-scale Widening Network (QAMWN) for endoscopy images.
• To address limitations of existing SR methods in terms of model size and channel-wise feature processing.

Main Methods:

• Developed QAMWN incorporating a stacked Quaternion Attention Multi-Scale Widening Block.
• Utilized Multi-scale Feature Widening Aggregation Module (MFWAM) for feature extraction.
• Introduced Quaternion Residual Channel Attention (QRCA) for adaptive feature scaling in the hyper-complex domain.

Main Results:

• QAMWN was evaluated on CVC ClinicDB and Kvasir endoscopic datasets.
• The proposed method achieved a favorable trade-off between model size and performance.
• QAMWN outperformed state-of-the-art methods in metrics and visualization.

Conclusions:

• A lightweight SR network (QAMWN) for endoscopy was proposed.
• The network achieves superior performance with fewer parameters, aiding clinical diagnosis.
• Quaternion attention mechanisms offer advantages in endoscopic SR.