PSD-ELGAN: A pseudo self-distillation based CycleGAN with enhanced local adversarial interaction for single image...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing CycleGAN-based dehazing methods struggle with accurately modeling haze-free features due to soft constraints.
• Global discriminators may misclassify hazy images with depth variations, leading to residual haze.

Purpose of the Study:

• To propose a novel image dehazing algorithm, PSD-ELGAN, that overcomes limitations of current methods.
• To enhance haze-free feature representation and improve residual haze suppression.

Main Methods:

• Utilizing CycleGAN for generating pseudo image pairs in an unsupervised framework.
• Employing pseudo self-distillation for knowledge transfer from clean to hazy image generators.
• Implementing enhanced local adversarial interaction with adaptive patch selection for residual haze removal.

Main Results:

• PSD-ELGAN demonstrates improved haze-free feature representation without increasing network parameters.
• The method effectively suppresses residual haze in challenging regions.
• Experimental results show promising performance and generality across diverse datasets.

Conclusions:

• PSD-ELGAN offers a robust solution for image dehazing by combining pseudo self-distillation and local adversarial learning.
• The proposed approach enhances the quality and clarity of de-hazed images.