Deep constrained multi-frame blind deconvolution with learned regularizations

Ren, Zhilei

doi:10.1364/AO.576616

Return

Deep constrained multi-frame blind deconvolution with learned regularizations

2025-12-10

PRE

DOI：10.1364/AO.576616

Original

Original request for help

Save

Abstract

En 中文

Multi-frame blind deconvolution (MFBD) is a widely used image restoration technique designed to compensate for atmospheric turbulence-induced wavefront distortions. It alternately estimates the original image and corresponding point spread functions (PSFs) from a sequence of short-exposure observed images with handcrafted regularizations based on limited prior information, which may lead to unstable results and are time-consuming in inference. To improve the efficiency and effectiveness, we propose an end-to-end trainable network for the space target image restoration task. In this network, the alternating estimations are performed on the deep features of the observed images using a constrained least-squares filter, where the smoothing filters are adaptively learned from the features. Following the common practices of MFBD, we design two convolutional neural modules for the PSFs estimation and deconvolution process, respectively. These two modules are cycled through repeatedly to recover a clear image. To train our network, a large-scale training dataset is constructed to enable its application in real turbulence scenarios. Extensive experiments show that the proposed method is robust and effective for high-resolution space target image restoration. (c) 2025 Optica Publishing Group. All rights, including for text and data mining (TDM), Artificial Intelligence (AI) training, and similar technologies, are reserved.

Keywords:

multi-frame blind deconvolution

deep learning

image restoration

point spread function estimation

adaptive regularization