An Attention-ReXNet Network for Long Tail Road Scene Classification

With the development of assisted driving and autonomous driving, road scene classification has become increasingly important due to its ability to provide environmental information for drivers and autonomous driving, which is beneficial to reduce traffic accidents. However, road scene classification is limited by dataset imbalance and running speed, making it difficult to apply practically. To solve the problems mentioned above, we proposed an Attention-ReXNet network to conduct road scene classification. A balance Softmax cross-entropy is introduced to solve the class imbalance problem in road scene classification. The proposed method is evaluated in the Road Surface Classification Dataset (RSCD), which contains 27 categories and 1 million images with class imbalance. Experimental results show that Attention-ReXNet and balance Softmax cross entropy can achieve the best classification performance, with accuracy and top-1 being 87.67% and 88.52%, respectively.