Automatic Bangla License Plate Recognition System for Low-Resolution Images

Automatic License Plate Recognition (ALPR) System is a crucial task that is used in numerous critical situations. Although there are several ways to accomplish LP recognition, our method is intended to be effective not just on LP with high resolution, but even when the LP is of very low resolution (LR). Previous works on ALPR systems emphasized achieving recognition benchmarks on LP images that are close to the sensor. However, if the camera is far from the vehicle, the generated detected frame yields an LP image of LR, which is hard for the recognition model to classify. Our method employed Enhanced Super Resolution Generative Adversarial Networks (ESRGANs) to upscale the [32 × 24] LR image to the high resolution of dimension [256 × 192] with a scale factor ratio of 8. For training purposes, we also collected and trained the models on a dataset of 2211 images. The Peak signal-to-noise ratio (PSNR) for the reconstruction on 200 testing samples is 12.587. The optical character recognition accuracy for the reconstructed and ground truth images are 78% and 91% respectively.