Multi-level Feature Interaction and Efficient Non-Local Information Enhanced Channel Attention for image dehazing

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image dehazing is a complex computer vision problem.
• Existing U-Net based methods struggle with feature utilization and information dilution.
• Standard Squeeze and Excitation (SE) channel attention can hinder performance in dehazing networks.

Purpose of the Study:

• To address limitations in current image dehazing techniques.
• To improve the recovery of edge details and overall scene aspects in dehazed images.
• To enhance the effectiveness of channel attention mechanisms in deep learning models for dehazing.

Main Methods:

• Proposed a novel Multi-level Feature Interaction and Non-local Information Enhanced Channel Attention (MFINEA) dehazing model.
• Introduced a multi-level feature interaction module for fusing shallow and deep features.
• Developed an efficient non-local information enhanced channel attention module for improved feature map weighting.

Main Results:

• The MFINEA model demonstrated superior performance compared to state-of-the-art methods.
• Enhanced recovery of edge details and overall scene quality in dehazed images.
• Validation on multiple challenging benchmark datasets confirmed the model's effectiveness.

Conclusions:

• The proposed MFINEA model effectively overcomes the limitations of existing dehazing methods.
• Multi-level feature interaction and enhanced channel attention significantly improve dehazing performance.
• MFINEA offers a promising advancement in computer vision for image restoration tasks.