Deep Retinex Network for Single Image Dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Hazy images degrade visual quality and hinder subsequent image analysis tasks.
• Existing dehazing methods often rely on simplified scattering models or prior knowledge, limiting their generalization ability.

Purpose of the Study:

• To propose a novel retinex-based decomposition model for hazy images.
• To develop an end-to-end deep retinex dehazing network (RDN) for effective image dehazing.

Main Methods:

• A retinex-based decomposition model separates hazy image illumination into natural and residual components.
• A deep retinex dehazing network (RDN) is designed, comprising a multiscale residual dense network for estimating the residual illumination map and a U-Net with attention mechanisms for dehazing.
• Channel and spatial attention mechanisms are integrated into the U-Net's skip connections to balance over- and underdehazing.

Main Results:

• The multiscale residual dense network effectively captures both global context and local details for precise residual illumination estimation.
• The attention-enhanced U-Net adaptively adjusts weights to optimize the dehazing process.
• The RDN demonstrates superior performance compared to scattering model-based, fully data-driven, and prior-based dehazing methods.

Conclusions:

• The proposed RDN offers improved generalization and avoids errors from simplified scattering models.
• The RDN achieves state-of-the-art performance in image dehazing tasks.