Vehicle Classification Enhancement Through Data Augmentation and Model Fusion: A Study of Deep Learning Approaches

In this paper, data enhancement and model fusion are used to improve the accuracy of vehicle classification. In terms of data sets, the experiment selected a dataset containing 1,600 images of ten types of vehicles and tested it using four deep learning models as a baseline, namely VGG16, ResNet50V2, InceptionV3, and DenseNet121. Then,the experiment used cutout technology to enhance the data of VGG16 and ResNet50V2, which had relatively poor performance in the test set, and achieved certain results. Then the experiment made model fusion of VGG16 and ResNet50V2. Model fusion was implemented at both feature level and probability level. At the feature level, fusion was performed after GlobalAveragePooling2D and Dense layer processing, including feature concatenation, addition, averaging, and maximum value selection. The fusion of probability level includes the weighted fusion of the output probabilities of the two models, taking the average value and the maximum value. The experimental results show that cutout data enhancement and model fusion have obvious effect on improving classification accuracy for the four baselines.