Neural Maximum a Posteriori Estimation on Unpaired Data for Motion Deblurring.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Real-world dynamic scene deblurring is challenging due to the lack of paired blurry-sharp training data.
• Conventional methods struggle with generalization due to handcrafted priors or synthetic data limitations.

Purpose of the Study:

• To propose a novel framework, Neural Maximum A Posteriori (NeurMAP), for training deblurring networks using unpaired data.
• To enable recovery of blind motion information and sharp content from real-world dynamic scenes.

Main Methods:

• Developed a joint training framework with a motion estimation network and a deblurring network.
• Modeled the (re)blurring process as a likelihood function.
• Utilized an implicit dynamic motion prior to guide deblurring network training.

Main Results:

• Demonstrated the first framework for training image deblurring networks on unpaired datasets.
• Achieved superior performance in both quantitative metrics and visual quality compared to state-of-the-art methods.
• Validated the effectiveness of the implicit dynamic motion prior.

Conclusions:

• NeurMAP offers an orthogonal approach to existing deblurring techniques.
• The proposed framework successfully addresses the challenge of training with unpaired data for dynamic scene deblurring.
• NeurMAP significantly advances the capabilities of real-world image deblurring.