BiN-Flow: Bidirectional Normalizing Flow for Robust Image Dehazing.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing image dehazing methods often rely on strict prior knowledge and paired training data, limiting their performance on unseen scenes.
• This dependency hinders the generalization capabilities of current dehazing models.

Purpose of the Study:

• To propose a novel image dehazing method, Bidirectional Normalizing Flow (BiN-Flow), that overcomes the limitations of prior knowledge and paired training.
• To improve the generalization and performance of image dehazing for diverse and unseen environmental conditions.

Main Methods:

• Developed Bidirectional Normalizing Flow (BiN-Flow) utilizing weakly-paired training, eliminating the need for strict prior knowledge.
• Introduced Feature Frequency Decoupling (FFD) using multi-scale residual blocks to capture intricate texture details.
• Implemented Bidirectional Propagation Flow (BPF) with invertible flows to model the complex relationships between hazy and haze-free images.
• Incorporated a reference mechanism (RM) leveraging both paired and unpaired haze-free reference images for robust training.

Main Results:

• BiN-Flow demonstrated superior performance and generalization capabilities compared to state-of-the-art methods across five standard datasets.
• The method effectively learned mutual relationships between hazy and haze-free images through weakly-paired training.
• Experimental results validated the capability of BiN-Flow in producing diverse dehazed images for a single input.

Conclusions:

• BiN-Flow offers a significant advancement in image dehazing by effectively addressing the limitations of traditional methods.
• The proposed approach achieves high performance and excellent generalization, making it suitable for real-world applications with varied conditions.
• The ability to generate diverse dehazed outputs adds a novel dimension to image restoration.