CVPR 2023 - Structured Kernel Estimation for Photon-Limited Deconvolution

In this episode we discuss Structured Kernel Estimation for Photon-Limited Deconvolution by Authors: Yash Sanghvi, Zhiyuan Mao, Stanley H. Chan Affiliation: School of Electrical and Computer Engineering, Purdue University. The paper proposes a new method for estimating blur in low light conditions with strong photon shot noise, where existing image restoration networks perform poorly. The authors use a gradient-based backpropagation method to estimate the blur kernel and model it using a low-dimensional representation with key points on the motion trajectory, reducing the search space and improving regularity of estimation. Results show improved performance compared to end-to-end trained neural networks when applied to deconvolution in an iterative framework. The code and pretrained models are available online.