DEA-Net: Single Image Dehazing Based on Detail-Enhanced Convolution and Content-Guided Attention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Single image dehazing is an ill-posed problem crucial for visual quality enhancement.
• Existing deep learning methods often increase model size for better performance.
• The learning capacity of Convolutional Neural Network (CNN) structures in dehazing remains an area for exploration.

Purpose of the Study:

• To propose a novel Detail-Enhanced Attention Block (DEAB) to boost feature learning for improved single image dehazing.
• To enhance the representation capacity and efficiency of dehazing models.
• To develop a Convolutional Neural Network (CNN) architecture that recovers high-quality haze-free images.

Main Methods:

• Introduced a Detail-Enhanced Attention Block (DEAB) comprising Detail-Enhanced Convolution (DEConv) and Content-Guided Attention (CGA).
• DEConv integrates prior information using difference convolutions and re-parameterization for efficiency.
• CGA utilizes Spatial Importance Maps (SIMs) for focused feature attention and employs a mixup fusion scheme to aid gradient flow.

Main Results:

• The proposed Detail-Enhanced Attention Network (DEA-Net) achieved state-of-the-art (SOTA) performance in single image dehazing.
• Achieved a Peak Signal-to-Noise Ratio (PSNR) exceeding 41 dB.
• Demonstrated effectiveness with a significantly reduced parameter count of only 3.653 million.

Conclusions:

• The DEAB, incorporating DEConv and CGA, effectively enhances feature learning for superior dehazing results.
• DEA-Net offers a computationally efficient and high-performing solution for single image dehazing.
• The proposed method outperforms existing SOTA techniques in terms of both performance and parameter efficiency.