Efficiency - Optimized Approach - Vehicle Classification Features Transfer Learning and Data Augmentation Utilizing Deep Convolutional Neural Networks

摘要

Vehicle Classification represents an essential function in the Traffic Management System. In recent years, predominantly, the Deep Convolutional Neural Network algorithms are widely adopted for object classification and detection. Accordingly, in this paper, transfer learning-based vehicle classification exercising pre-trained Deep Convolutional models such as VGG16, InceptionV3 are proposed. To reduce the over-fitting problem of Deep Convolutional Neural Networks, on minimum-capacity datasets Transfer Learning and Data Augmentation methods are enabled in this proposed system. The performance of the model is tested, premised on the experiments on the custom dataset of vehicle images. In this study, the classification and detection algorithm seeks three classes of vehicles such as bus, truck, and motorcycle. Consequently, the experimental outcomes reveal that compared with the VGG16 model, the classification accuracy of the pretrained model is higher by implementation of the InceptionV3 model. The InceptionV3 model with an optimized approach achieves classification accuracy of 99.33% for the training set and 98.87% for the validation set, which governs improvement in accuracy of detection.