Deep Dehazing Network With Latent Ensembling Architecture and Adversarial Learning

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing image dehazing algorithms often rely on estimated transmission maps and atmospheric light, leading to unrealistic results due to inaccurate estimations and model assumptions.
• Traditional methods struggle with uniform contrast enhancement and fail to capture the spatial non-uniformity of haze effectively.

Purpose of the Study:

• To develop an end-to-end, photo-realistic image dehazing algorithm using an adversarial game between neural networks.
• To address the limitations of existing methods by simultaneously restoring haze-free images and capturing haze non-uniformity.
• To improve the adaptability of dehazing techniques for high-level computer vision tasks like object detection.

Main Methods:

• Employed an adversarial game between a generator and a multi-scale discriminator for end-to-end photo-realistic dehazing.
• Designed a generator architecture with sequential and parallel modules for information sharing, implicitly forming an ensemble of dehazing models.
• Utilized identity mapping in clear-scene image space for data-driven regularization, avoiding hand-crafted loss functions for artifact penalization.
• Introduced a task-driven training strategy to optimize object detection performance on dehazed images without modifying the object detector.

Main Results:

• The proposed algorithm achieves superior performance in realistic image dehazing, outperforming ten state-of-the-art methods on RESIDE, I-Haze, and O-Haze benchmarks.
• Demonstrated effective capture of non-uniform haze and reduction of dehazing artifacts.
• Successfully improved object detection performance on dehazed images through the task-driven training strategy.

Conclusions:

• The adversarial game approach offers a robust and effective solution for photo-realistic image dehazing.
• The proposed method overcomes limitations of traditional algorithms by handling haze non-uniformity and providing adaptable features for downstream tasks.
• This work advances the field of image dehazing and its application in computer vision.