Self-Parameter Distillation Dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image dehazing is crucial for improving visual quality in adverse weather conditions.
• Conventional knowledge distillation requires separate teacher and student networks.
• Existing methods often struggle with complex haze patterns and information loss.

Purpose of the Study:

• To propose a novel and efficient image dehazing method using self-distillation.
• To develop a single network capable of transferring knowledge internally for improved dehazing.
• To enhance the accuracy and performance of image dehazing algorithms.

Main Methods:

• A single knowledge distillation network architecture is proposed.
• The network utilizes self-distillation by transferring parameters to itself in stages.
• Early stages transfer scene content (identity) using haze-free data; later stages transfer haze information using haze data.
• Forward propagation acts as the teacher, and backward propagation acts as the student during training.

Main Results:

• The proposed self-distillation method achieves accurate dehazing by leveraging early-stage scene information in later stages.
• Experimental results show the method significantly outperforms state-of-the-art dehazing techniques.
• The single network architecture allows seamless parameter updates from feature extraction to dehazing.

Conclusions:

• Self-distillation offers an effective approach for single-network image dehazing.
• The proposed method provides a computationally efficient and high-performance solution for image restoration.
• This work advances the field of image processing by introducing a novel distillation paradigm.