Enhancing image restoration through learning context-rich and detail-accurate features

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image restoration seeks high-quality images from degraded ones, balancing spatial details and context.
• Existing methods often neglect frequency variations and can introduce noise via direct feature fusion in skip connections.

Purpose of the Study:

• To develop an image restoration model that optimally balances spatial and frequency domain information.
• To mitigate noise propagation through skip connections in deep learning architectures for image restoration.

Main Methods:

• Introduced a hybrid scale frequency selection block (HSFSBlock) for multi-scale spatial and frequency domain analysis.
• Developed a skip connection attention mechanism (SCAM) to selectively filter information passed through skip connections.
• Proposed a tightly interlinked architecture named LCDNet.

Main Results:

• LCDNet effectively integrates spatial and frequency domain knowledge for selective information recovery.
• The SCAM mechanism successfully mitigates noise introduced by conventional skip connections.
• Experimental results demonstrate superior or comparable performance to state-of-the-art algorithms across various image restoration tasks.

Conclusions:

• LCDNet offers an effective approach to image restoration by addressing limitations in existing methods.
• The proposed HSFSBlock and SCAM contribute to improved performance and noise reduction.
• The model shows strong potential for diverse image restoration applications.