Abstract

Automatic License Plate Detection and Recognition (ALPR) has many practical applications such as traffic control and parking tickets; for this reason, it has been one of the exciting research topics. Environmental factors such as lighting and dust, make automatic license plate detection and recognition challenging, especially for traditional image processing methods. Although much research has been conducted on ALPR systems using image processing and computer vision tools and algorithms, the need for more research on this topic with deep-learning algorithms has not been satisfied yet. Among different and in succession phases of ALPR, the license plate detection phase is of great importance because it is the first phase, and its performance affects the result of other stages. Moreover, due to the advent of technology and artificial intelligence in everyday life, having reliable real-time ALPR systems is necessary. Hence, this work empirically studies the mean Average Precision (mAP) of Single Shot MultiBox Detector (SSD) and You Only Look Once (YOLOv4) on CENPARMI and UFPR-ALPR datasets. Although we achieved good mAP results of 95.47 % (ResNet-SSD) and 95.45 % (InceptionV2-SSD) with the SSD model during this experiment, we have reached the highest mAP of 97.46 % and 97.78 % with the newly released YOLOv4 model on CENPARMI and UFPR-ALPR datasets, respectively. However, in object detection, high precision is not the only essential criterion anymore. Hence, we scrutinized the object-detectors mentioned above to find a model that can balance mAP, speed, and memory. We learned that the higher the number of parameters of a model, the better the detection results. On the other hand, the number of parameters of a model can affect an object detection task’s speed.