CNN for License Plate Motion Deblurring

TL;DR

This paper demonstrates that convolutional neural networks trained on artificially generated data can effectively deblur real traffic surveillance images, outperforming traditional methods especially when tailored to specific blur characteristics.

Contribution

It introduces a method for training CNNs on synthetic data for license plate deblurring, showing superior results on real images and analyzing the limits of CNN performance based on blur parameters.

Findings

CNNs trained on artificial data outperform traditional methods on real images.

Training data can be generated easily from sharp images with approximate blur kernels.

The study evaluates CNN performance limits based on blur direction and length.

Abstract

In this work we explore the previously proposed approach of direct blind deconvolution and denoising with convolutional neural networks in a situation where the blur kernels are partially constrained. We focus on blurred images from a real-life traffic surveillance system, on which we, for the first time, demonstrate that neural networks trained on artificial data provide superior reconstruction quality on real images compared to traditional blind deconvolution methods. The training data is easy to obtain by blurring sharp photos from a target system with a very rough approximation of the expected blur kernels, thereby allowing custom CNNs to be trained for a specific application (image content and blur range). Additionally, we evaluate the behavior and limits of the CNNs with respect to blur direction range and length.