Hierarchical Joint CNN-Based Models for Fine-Grained Cars Recognition

For the purpose of public security, car detection and identification are urgently required in the real time traffic monitoring system. However, fine-grained recognition is a challenging task in the area of computer vision due to the subtle inter-class and huge intra-class differences. To tackle this task, this paper provided a novel approach focussed on two main aspects. On the one hand, the most discriminative local feature representations of regions of interests (ROIs) magnified many details. On the other hand, the hierarchical relations within the fine-grained categories can be simulated by probability formulas. Our proposed model consists of two modules: (i) a region proposal network to generate plenty of ROIs and (ii) a joint CNN-based model to learn the multi-grained feature representations simultaneously. The proposed joint CNN-based model was implemented and tested on the Stanford Cars dataset and the CompCars dataset. Our experimental results are compared with those of other methods, and verify the superior performance of the proposed model.