Abstract: Most existing fine-grained car recognizers have achieved the good performances on various datasets. However, these recognizers usually work inefficiently, and attention regions cannot be accurately localized especially when multiple cars exist. In this work, we propose a fast and accurate car recognizer called FACR, and it has three advantages over the previous approaches as follows. 1) The attention region localization component proposed in this work has high robustness to the background noise in the image. 2) The network structure proposed in this work is relatively simple, and it has higher computational efficiency while maintaining near accuracy. 3) It is the first time to use the hierarchical classification to finely and accurately categorize cars, and to allow a single network to simultaneously recognize car types, manufacturers, car models, and years of manufacture. The experiments are carried out on standard datasets Compcars and Stanford Cars, and the results demonstrate that FACR achieves the competitive performance compared with the-state-of-art methods in terms of both computational efficiency and classification accuracy.

Key words: fine-grained recognition, car recognition, deep learning