Dynamic High-Pass Filtering and Multi-Spectral Attention for Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) have advanced super-resolution (SR) but are biased towards low-frequency signals.
• This bias hinders the reconstruction of fine image details and textures crucial for SR.
• Existing SR models struggle to effectively capture and reconstruct high-frequency information.

Purpose of the Study:

• To address the limitations of CNNs in learning high-frequency features for image super-resolution.
• To introduce novel architectural modules that improve the preservation and reconstruction of high-frequency details.
• To enhance the accuracy and visual quality of super-resolved images.

Main Methods:

• Proposed a dynamic highpass filtering (HPF) module for adaptive, local preservation of high-frequency signals.
• Introduced a matrix multi-spectral channel attention (MMCA) module for global recalibration of frequency-domain features.
• Integrated these modules into existing SR architectures to improve high-frequency feature learning.

Main Results:

• The proposed HPF and MMCA modules demonstrated significant improvements in learning high-frequency features.
• Experiments showed superior accuracy and visual quality compared to state-of-the-art SR methods.
• The modules effectively preserved and reconstructed fine image textures and details.

Conclusions:

• The novel HPF and MMCA modules effectively mitigate the low-frequency bias in CNN-based SR.
• These modules offer a promising approach for enhancing high-frequency feature learning in image super-resolution.
• The proposed methods achieve state-of-the-art performance on benchmark datasets.