A Lightweight Fusion Distillation Network for Image Deblurring and Deraining

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning methods for image deblurring and deraining are advanced but often struggle with effective feature distillation.
• Existing approaches require numerous parameters, leading to high computational costs.

Purpose of the Study:

• To propose a lightweight fusion distillation network (LFDN) for efficient image deblurring and deraining.
• To address the limitations of feature distillation and high computational burden in current deep learning models.

Main Methods:

• An encoder-decoder architecture is employed for multi-scale information extraction and fusion.
• A feature distillation normalization block is introduced to continuously screen valuable channel information.
• An attention mechanism facilitates information fusion between distillation modules and feature channels.

Main Results:

• The proposed LFDN achieves state-of-the-art performance in image deblurring and deraining.
• The network demonstrates superior results with a significantly smaller number of parameters compared to existing methods.
• The LFDN outperforms previous approaches in terms of model complexity and computational efficiency.

Conclusions:

• The lightweight fusion distillation network (LFDN) effectively addresses challenges in image deblurring and deraining.
• The LFDN offers a computationally efficient solution with state-of-the-art performance.
• This approach advances the field by enabling high-quality image restoration with reduced model complexity.