Random Multi-Augmentation to Improve TensorFlow-Based Vehicle Plate Detection

In the development of the "Machine Learning" education kit, vehicle plate recognition was created using TensorFlow with SSD MobileNetV2. The detection failure rate in the training process with varying distances and lighting from the camera is high if the training data is insufficient. Addressing that notable gap in research, we proposed Random Multi-Augmentation to Improve TensorFlow-Based Vehicle Plate Detection. Augmentation techniques are expected to train data that is manipulated at varying lighting and distance. The proposed method consists of two combining augmentation approaches, namely: position augmentation and lighting augmentation. Position augmentation which consists of Flip, Crop, Rotate, Shift, and Crop is used to enrich the visualization of distance and viewing angle, while Lighting augmentation which consists of Greyscale, Hue, Saturation, Brightness, Exposure, and Blur is used to enrich the visualization of lighting. Variations in values were determined randomly based on variations in values from several previous studies. The comparison of TensorFlow SSD MobileNetV2 and Augmentation were tested using one video Roboflow. TensorFlow without augmentation exhibited an accuracy of 60%, precision of 100%, recall of 60%, and an F1 score of 75%, whereas TensorFlow within augmentation achieved a higher accuracy of 70%, precision of 100%, recall of 70%, and an F1 score of 82.3%. Based on precision measurement, Tensorflow can be claimed to prevent false positives, which indicates that the algorithm did not detect non-plate objects as vehicle plates. Furthermore, a comparison of the use of augmentation shows an increase in plate detection capabilities when using augmentation as Tensorflow preprocessing, which is indicated by an increase in recall and accuracy values. These results emphasize that augmentation is the pre-processing optimizer for vehicle number plate systems.