Deep Learning based Radial Blur Estimation and Image Enhancement

Abstract

In this paper, we propose a deep learning based pipeline to estimate the radial blur and enhance the deblurred image. The radial blur is introduced in the image as an effect of ego motion in autonomous vehicle systems. The deblurring of the image with radial blur is challenging since most of the blur models do not estimate radial blur. Hence, we design a deep learning based pipeline with estimation and enhancement modules. The estimation module is designed with CuratorNet to estimate radial PSF in two stages. The estimated PSF is used for deblurring of input radial blurred images. The enhancement module is designed with convolutional autoencoder which enhances the deblurred image to remove artefacts in order to detect the traffic signs. We demonstrate the results of the proposed pipeline on synthetic and real images with traffic signs and compare the results with existing methods.