Improved architecture for traffic sign recognition using a self-regularized activation function: SigmaH

Traffic sign recognition (TSR) is a crucial intelligent transport system. Nowadays, the convolutional neural network has become a vital tool in the conception of a TSR model. In this work, we propose an improved TSR algorithm for the transportation system inspired by the classical model LeNet-5. In this model, firstly, we replace the hyperbolic tangent activation function with a self-regularized non-monotonic activation function called SigmaH $$({\text{SigmaH}}(x) = x\tanh (\sqrt {\sigma \left( x \right)} ))$$ . SigmaH is experimentally validated on various popular benchmarks against the most suitable combinations of architectures and activation functions. Secondly, we use a convolutional block attention module, which is beneficial for extracting the most valuable features using the attention method. Finally, we combine the triplet-center loss with the Softmax activation function as a loss function to maximize the correct recognition rate. The TSR experiments are carried out based on the German Traffic Sign Recognition benchmark. The experimental results demonstrate that the improved LeNet-5 has an identification accuracy rate of 98.25%, and the average processing time per frame is 8 ms. In the meantime, the number of parameters is reduced by more than 51% compared with the classic LeNet-5 model. Our proposed model has remarkable accuracy and high training efficiency compared with other algorithms.