Abstract:

The implementation of autonomous driving technology required high-precision recognition of traffic signs. However， due to their high similarity， small size， and vulnerability to outdoor environmental factors， achieving real-time and accurate detection posed significant challenges. In response to the limitations of traditional neural network design approaches， an algorithm based on self-selecting architecture was proposed to automatically adjust the network structure， thereby enhancing model performance and efficiency. The algorithm adopted a two-stage training approach to optimize the selection of network paths. Moreover， gradient propagation was employed to train the hyperparameters for multiple loss functions， replacing the conventional manual tuning with a dynamic loss network scheme. The results demonstrated that the proposed algorithm achieved an accuracy rate of 95.74% and a detection speed of 146.58 frames per second on the GTSRB dataset， while maintaining a model parameter size of only 0.46Mb， enabling deployment on mobile devices. Compared to the traditional manual design of static networks， the adoption of the self-learning architecture module not only reduced experimental costs but also improved accuracy and performance. Furthermore， it enabled superior detection outcomes in various environments and exhibited a noticeable enhancement in loss convergence speed.

Key words: traffic sign classification, deep learning, self-selecting architecture, optimal network path, dynamic loss network