Task-Oriented Network for Image Dehazing | JoVE Visualize

Area of Science:

Computer Vision
Image Processing
Artificial Intelligence

Background:

Haze significantly degrades outdoor image quality by interfering with radiation transmission, affecting color and detail.
Current deep neural network (DNN) based dehazing algorithms often employ generic networks, failing to accurately model the haze formation process.
This leads to artifacts and residual haze in dehazed images, especially in challenging scenarios.

Purpose of the Study:

To propose a novel task-oriented network for image dehazing that accurately models the physical process of haze formation.
To develop a multi-stage dehazing algorithm for enhanced accuracy and removal of haze residuals.
To introduce effective loss functions and constraints for robust network training.

Main Methods:

A hybrid network architecture combining an encoder-decoder structure with a spatially variant recurrent neural network (RNN) derived from the hazy process.
A multi-stage dehazing algorithm designed for step-by-step filtering of haze residuals.
Development of a dual composition loss, content-based pixel-wise loss, and total variation constraint for network regularization.
End-to-end training and analysis of the proposed network's performance.

Main Results:

The proposed task-oriented network effectively models the image formation of haze, leading to improved dehazing performance.
The multi-stage algorithm successfully filters out haze residuals, enhancing image clarity and detail.
Experimental results show the algorithm achieves favorable performance compared to state-of-the-art dehazing methods.
Dehazed images exhibit fewer artifacts and reduced haze residuals, even in special scenes.

Conclusions:

The proposed task-oriented network, motivated by the physical haze formation process, offers a significant advancement in image dehazing.
The hybrid network architecture and multi-stage refinement effectively address limitations of existing methods.
The developed loss functions and constraints contribute to robust and accurate dehazing, outperforming current state-of-the-art techniques.