Deep-blur: Blind identification and deblurring with convolutional neural networks.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Image deblurring from a single degraded image is a challenging problem in computer vision.
• Estimating the blurring operator (point spread function) is crucial for effective deblurring.
• Existing methods often struggle with noise and complex, space-varying blurring operators.

Purpose of the Study:

• To develop a novel neural network architecture for estimating blurring operators.
• To enable robust image deblurring using the estimated operator from a single image.
• To improve deblurring performance compared to current state-of-the-art methods.

Main Methods:

• Proposed a neural network architecture and training procedure to estimate blurring operators.
• Parameterized forward operators using low-dimensional vectors, including Zernike polynomials and product-convolution expansions for space-varying operators.
• Utilized the estimated operator as input for an unrolled neural network for image deblurring.

Main Results:

• Accurate and robust recovery of blur parameters demonstrated, even under high noise levels.
• Achieved signal-to-noise ratios for recovered point spread functions ranging from 13 dB (noiseless) to 8 dB (high noise).
• Outperformed alternative methods perceptually and in terms of Structural Similarity Index Measure (SSIM).

Conclusions:

• The proposed method effectively estimates blurring operators and deblurs images from a single degraded input.
• The approach demonstrates superior performance and robustness against noise compared to existing techniques.
• The algorithm offers fast processing speeds on consumer hardware without requiring human interaction post-setup.