Self-supervised zero-shot dehazing network based on dark channel prior.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Supervised learning for image dehazing demands extensive datasets and significant training time.
• Acquiring large-scale, high-quality datasets for image dehazing is a major challenge.
• Existing methods often struggle with accuracy and efficiency due to data limitations.

Purpose of the Study:

• To develop a self-supervised zero-shot dehazing network (SZDNet) that eliminates the need for large training datasets.
• To improve the accuracy of atmospheric light estimation in hazy images.
• To enhance the quality of dehazed images through a novel loss function.

Main Methods:

• Proposed a self-supervised zero-supervised dehazing network (SZDNet) utilizing the dark channel prior.
• Implemented a novel multichannel quad-tree algorithm for accurate atmospheric light value estimation.
• Employed a combined cosine distance and mean squared error loss function for optimization.

Main Results:

• SZDNet successfully dehazes images without requiring pre-training on large datasets.
• The multichannel quad-tree algorithm provides more accurate atmospheric light estimations.
• The proposed loss function enhances the visual and quantitative quality of dehazed images.
• Achieved competitive performance against state-of-the-art dehazing methods.

Conclusions:

• SZDNet offers an efficient and effective solution for image dehazing, overcoming data dependency.
• The method demonstrates significant potential for real-world applications requiring rapid image enhancement.
• The self-supervised approach represents a promising direction for future research in image restoration.