Self-supervised Defocus Map Estimation and Auxiliary Image Deblurring Given a Single Defocused Image.

In this paper, we propose an end-to-end self-supervised Deep Neural Network (DNN) for Defocus Map Estimation (DME). Currently, such defocus maps are estimated by DNNs with fully supervised learning. For training, such networks need large datasets annotated by defocus amount or scene depth, which are challenging to obtain. The networks with self-supervised training arrive to overcome the limitation represented by the Ground Truth (GT) data. In this line, we propose a self-supervised learning neural network for DME from a single defocused image. Our method is based on a recently proposed DNN called 2HDED:NET that we enriched with a defocus simulation module that makes possible the self-supervised training for DME. In addition to the defocus map, our network reconstructs the All-in-Focus (AIF) image through supervised learning. We test the network on synthetic and realistic benchmarks and demonstrate that it is an effective solution for DME and image deblurring when a single defocused image is available.