Super-resolution of license plates

Abstract

Nowadays, image enhancement plays an essential role in crime scene investigation and the monitoring of offences on public roads. CCTV videos usually contain several distortions, including poor illumination, low-resolution, optical distortion and compression artefacts. As a consequence, specific important details may be blurred, hindering recognition of license plates. Super-resolution can be mainly divided into two categories: Single-image super-resolution and Multiple-image super-resolution. Currently, multiple-image super-resolution is the main technique which is being used to improve image quality in forensic science. However, this approach has various limitations. In this dissertation, single-image super-resolution techniques were adapted to restore license plates. This project explores the use of class-based super-resolution techniques which were successfully implemented in face and iris super-resolution. This work proposes three different approaches to class-based single-image super-resolution. The first methodology which was used is the simple least-squares method. This technique consisted of posing the super-resolution problem as a least-squares problem to minimise the reconstruction error of the images and achieve appropriate reconstruction weights. The second approach was super-resolution by Eigen transformation. This method utilises principal component analysis (PCA) to perform the super-resolution procedure while significantly reducing the dimensionality of the dictionaries which were used. Finally, a neighbour embedding method was implemented. This method uses a patch-based approach, which divides the images into a number of overlapping patches and performs the super-resolution procedure on every patch in order to obtain the best reconstruction weights for each patch. Experiments show that the first two methods presented overfitting which significantly hinders the character recognition process due to the global nature of these approaches. Thus, they are not able to adapt to different license plate structures. On the other hand, the neighbour embedding method yielded satisfactory results. This method was consistently shown to surpass the perceived quality of license plates processed through bilinear interpolation.