Application Optimization of Fine-grained Vehicle Classification based on Backbone Network

Abstract

As an important branch of fine-grained visual classification, relevant research on fine-grained vehicle model classification is of great significance to the promotion of smart travel and transportation applications. In this paper, we propose a set of optimization schemes for fine-grained vehicle model classification based on the backbone network like EfficientNet. The presented schemes, which have been relatively few studied in the other networks, is optimized through input data, training and inference process with little extra costs but distinct benefit. In particular, based on the characteristics of the vehicle model images, a novel data augmentation method, Center Fill with Shifted Quadrants (CFSQ), is proposed to facilitate effective feature extraction during the training and inference stage. Experiments on Stanford Car-196 dataset validate the accuracy and efficiency of our proposed method, which could achieve a high classification accuracy close to the state-of-the-art methods. Moreover, reasonable suggestions for the improvement and application are provided of the Stanford Car-196 dataset in this paper.