Traffic signs recognition using dynamic-scale CNN

Qiang Chen, Shaoyu Wang, Lirong Yi, Wenting Shu, Rui Wang, Jue Chen

Abstract: This paper presents an improved CNN-based structure to recognize the traffic signs. In the conventional CNN structure, only features extracted by the last convolutional layer are fed into the first fully connected layer. A recently proposed method for using more features, the MS-CNN uses features extracted by all convolutional layers, but this method may ignore the differences among layers. In our improved structure, the previous convolutional feature maps are combined to the first fully connected layer by a dynamic priority algorithm, which is inspired by the dynamic priority scheduling in the operating system. Our improved structure assigns different priority weights for certain convolutional layers, and priority is dynamically changing. This structure emphases differences among various convolutional layers and trains the network from different scales of features with assigned priorities. The results on the GTSRB dataset show that our method achieves recognition accuracy up to 97.61%, which is a considerable improvement compared to the conventional CNN and the MS-CNN.