Estimation of non-uniform motion blur using a patch-based regression convolutional neural network (CNN)

• Luis Varela, Laura E Boucheron, Steven Sandoval, David Voelz, and Abu Bucker Siddik
• received 02/06/2024; accepted 04/18/2024; posted 04/19/2024; Doc. ID 521076
• Abstract: The non-uniform blur of atmospheric turbulence can be modeled as a superpositionof linear motion blur kernels at a patch level. We propose a regression convolutional neuralnetwork (CNN) to predict angle and length of a linear motion blur kernel for varying sizedpatches. We analyze the robustness of the network for different patch sizes and the performanceof the network in regions where the characteristics of the blur are transitioning. Alternating patchsizes per epoch in training, we find coefficient of determination scores across a range of patchsizes of 𝑅2 > 0.78 for length and 𝑅2> 0.94 for angle prediction. We find that blur predictions inregions overlapping two blur characteristics transition between the two characteristics as overlapchanges. These results validate the use of such a network for prediction of non-uniform blurcharacteristics at a patch level.