Traffic Sign Recognition Algorithm Based on Multi-scale Convolution and Weighted-Hybird Loss Function
Zhengjie Lan, Li-e Wang, Zhiming Su · 2021
Traffic sign recognition is an important component of intelligent transportation system, it is a rather challenging work due to its real-time and high accuracy requirements. Therefore, this paper proposes a lightweight model based on Multi-scale Depth Separable convolution. In order to reduce the parameters, we designed a Multi-scale Asymmetric Convolution block and embed the Depthwise-Separable convlution into the Inception-A structure. Meanwhile, this study aims to improve the ability of feature extraction for hard-to-classify samples in the imbalance dataset, a Weighted-Hybrid loss function is presented to improve the model's focus more on the characteristics of the samples hard-to-classify. Finally, the model achieves 99.33% and 98.92% recognition accuracy on Belgium TS dataset and GTSRB dataset respectively. In the meantime the amount of network parameters and Floating Point Pperations (FLOPs) decreased significantly.