Research and Implementation of License Plate Location Based on Improved YOLO Algorithm

Abstract

At present, license plate location technology mainly has traditional methods and deep learning methods. The traditional license plate location mainly relies on manual extraction of features, and according to the shape and proportion of the license plate to achieve license plate positioning. However, the traditional license plate location method is difficult to meet the requirements of license plate size, shape, illumination and other changes, and it is difficult to achieve the robustness of license plate location. The license plate location technology based on deep learning does not need to select features and design classifiers manually. Its implementation is mainly based on the network automatically learning features of the data to achieve end-to-end license plate location. In this paper, we mainly study You Only Look Once (YOLO) algorithm, which is not accurate enough for the small target positioning, and in order to further improve the license plate location accuracy and speed, we propose to use the large feature map size of 14x14, improve the loss function by adding the Generalized Intersection over Union (GIoU) loss, and when it is training, the Intersection over Union (IoU) calculation method is changed to GIoU, the feature extraction network uses ShuffleNetV2, MobileNetV2 and other networks to extract more accurate features. In the data about campus bayonet vehicles, compared with Single Shot MultiBox Detector (SSD) and YOLOv2, the improved YOLO algorithm has better location accuracy and less location time. The Mean Average Precision (mAP) of improved DetNet59 have reached 99.79% and the precision (prec) has reached 99.97%. For the color image with a width of 680 pixels, and height of 480 pixels, the license plate location time is about 47 milliseconds (ms).