Estimation of non-uniform motion blur using a patch-based regression convolutional neural network.

Area of Science:

• Computer Vision
• Image Processing
• Astrophysics

Background:

• Atmospheric turbulence causes non-uniform image blur.
• This blur can be modeled as a combination of linear motion blur kernels at a patch level.

Purpose of the Study:

• To develop a regression convolutional neural network (CNN) for predicting linear motion blur kernel parameters (angle and length).
• To analyze the network's robustness across different patch sizes and its performance in transitional blur regions.

Main Methods:

• A regression CNN was designed to predict blur angle and length.
• The network was trained using alternating patch sizes per epoch.
• Performance was evaluated across various patch sizes and in areas with transitioning blur characteristics.

Main Results:

• High prediction accuracy was achieved, with R² scores above 0.78 for length and 0.94 for angle across a range of patch sizes.
• Blur predictions in overlapping regions smoothly transitioned between blur characteristics.
• The network demonstrated robustness for varying patch sizes.

Conclusions:

• The proposed CNN effectively predicts non-uniform blur characteristics at a patch level.
• The findings validate the use of CNNs for analyzing complex atmospheric turbulence effects in images.