Abstract

A vehicle registration plate is used for vehicle identity. In recent years, technology to identify plate numbers automatically or known as Automatic License Plate Recognition (ALPR) has grown over time. Convolutional Neural Network and YOLACT are used to do plate number recognition from a video. The number plate recognition process consists of 3 stages. The first stage determines the coordinates of the number plate area on a video frame using YOLACT. The second stage is to separate each character inside the plat number using morphological operations, horizontal projection, and topological structural. The third stage is recognizing each character candidate using CNN MobileNetV2. To reduce computation time by only take several frames in the video, frame sampling is performed. This experiment study uses frame sampling, YOLACT epoch, MobileNet V2 epoch, and the ratio of validation data as parameters. The best results are with 250ms frame sampling succeed to reduce computational times up to 78%, whereas the accuracy is affected by the MobileNetV2 model with 100 epoch and ratio of split data validation 0,1 which results in 83,33% in average accuracy. Frame sampling can reduce computational time however higher frame sampling value causes the system fails to obtain plate region area.