Publication

Abstract : Automatic License Plate Recognition (ALPR) is a technology that recognizes license plates from a given image and then uses optical character recognition on it to read the registration plates. We propose a two-staged approach to accomplish this. In the first stage, we use transfer learning with state-of-the-art YOLO object detector. This gives an efficient algorithm to detect the exact location of the license plates in the given image. We compare two versions of YOLO (YOLOv2 and YOLOv3) to gain insight into the differences between the versions. Then, in the second stage, we use image enhancement techniques to run optical character recognition to obtain the final output text. The system is trained on a custom dataset, in which we have manually annotated all the images. We use Intersection over Union (IoU) and mean Average Precision (mAP) as our evaluation metrics and obtain an IoU of 58.24% and mAP of 53.30% for YOLOv3 and 57.91% IoU and 51.28% mAP for YOLOv2.