A Multi-Scale Dehazing Network with Dark Channel Priors.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional neural networks (CNNs) show promise in image dehazing.
• Existing CNN-based methods struggle with incomplete dehazing, color deviation, and loss of detail.

Purpose of the Study:

• To propose a novel multi-scale dehazing network with dark channel priors (MSDN-DCP).
• To enhance feature extraction, fusion, and refinement for superior image dehazing.

Main Methods:

• Introduced a feature extraction module (FEM) with a two-branch residual structure.
• Devised a feature fusion module (FFM) for adaptive multi-scale feature combination.
• Proposed a dark channel refinement module (DCRM) using dark channel prior theory.

Main Results:

• Achieved a peak signal-to-noise ratio (PSNR) of 29.57 dB on the Haze4K dataset.
• Attained a structural similarity (SSIM) of 98.1% on the Haze4K dataset.
• Demonstrated superior dehazing performance over existing algorithms in objective metrics and visual perception.

Conclusions:

• The proposed MSDN-DCP effectively overcomes limitations of previous CNN-based dehazing methods.
• The network achieves state-of-the-art results in image dehazing.
• MSDN-DCP offers improved visual quality and detail preservation in dehazed images.