Efficient Dehazing with Recursive Gated Convolution in U-Net: A Novel Approach for Image Dehazing

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image dehazing is crucial for recovering clear visuals from hazy images.
• Deep learning has advanced image dehazing, but complex networks hinder practical application.
• Existing methods often involve intricate architectures, complicating training and deployment.

Purpose of the Study:

• To propose an efficient and effective end-to-end U-Net based image dehazing network.
• To maintain a lean network structure while enhancing dehazing performance.
• To introduce a novel U-Net variant named Dehaze Recursive Gated U-Net (DRGNet).

Main Methods:

• Developed an end-to-end U-Net dehazing network incorporating recursive gated convolution and attention mechanisms.
• Replaced standard convolution blocks with residual blocks using an improved recursive gated convolution.
• Utilized the SK fusion module to enhance the skip connection methodology.

Main Results:

• DRGNet demonstrated superior performance in dehazing quality and detail retrieval on public datasets.
• Objective evaluation metrics confirmed the effectiveness of the proposed model.
• Ablation studies validated the contribution of key design elements within DRGNet.

Conclusions:

• The proposed DRGNet offers a computationally efficient and high-performing solution for image dehazing.
• The integration of recursive gated convolution and attention mechanisms in a U-Net architecture is effective.
• DRGNet presents a practical advancement in computer vision for clear image recovery.