Structure-Aware Motion Deblurring Using Multi-Adversarial Optimized CycleGAN.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) show promise in blind image motion deblurring.
• Existing methods often require paired data and struggle with preserving structural information.

Purpose of the Study:

• To develop an unsupervised image deblurring method that overcomes limitations of existing approaches.
• To enhance structural information retention in deblurred images.

Main Methods:

• Utilizing a multi-adversarial optimized cycle-consistent generative adversarial network (CycleGAN).
• Implementing iterative high-resolution image generation with gradual supervision of generator hidden layers.
• Introducing a structure-aware mechanism with edge map guidance and multi-scale edge constraints.

Main Results:

• The proposed method successfully performs blind motion deblurring without paired training data.
• It significantly improves the retention of structural information in deblurred images.
• Experimental results demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The multi-adversarial CycleGAN with a structure-aware mechanism offers an effective solution for unsupervised blind image motion deblurring.
• This approach eliminates the need for paired data and blur kernel estimation, enhancing practical applicability.
• The method achieves superior performance in maintaining image structure and details.