License Plate Detector Application with Single Shot Detector Model

Abstract

Convolutional Neural Networks are one the best technologies used in license plate detection, specifically Single Shot Detector (SSD), which has great speed and accuracy. There is a lack of studies on the effects between epochs of batch size towards time and mAP in SSD research. Hence, a license plate detector using the SSD model is developed, and the optimal batch size and hyperparameters are identified. A dataset containing 7058 training images, 2048 validation images and 1020 testing images are used. We obtain the highest mAP of 66.13% when the batch size is 2. Generally, when batch size increases, the computation time and mAP decreases. An epoch of 2 has the lowest training time of 0.39 hours but trades off mAP of only 62.09%. The mAP gradually increases when the epoch increases to 8 epochs at 65.53% and stagnates at higher epochs. Thus, the optimal hyperparameters are batch size of 2 and 8 epochs. Then, an application based on the license plate detector is created. Using a small sample of images obtained from the Universiti Tun Hussein Onn campus, the application can detect completely grayscale images unaffected by the lack of colour. The application can best detect front license plates at 4 meters and rear license plates at 1 meter with confidence of 91% and 97%, respectively. The best angle detected has a confidence score of 99%.