Traffic Sign Recognition Algorithm Based on Sparse Coding End-to-End Training

This study aims at improving traffic sign recognition accuracy at different angles of view and the recognition accuracy of the existing algorithms applied in identifying traffic signs with angular deformations. For this purpose, we put forward an end-to-end traffic sign recognition approach based on sparse coding. Coordinate relations of various features could be reserved depending on a CapsNet feature extraction layer that carries out joint discrimination by virtue of vector length and direction. In this way, the network's ability to discriminate samples with similar shapes was boosted on one hand; and on the other hand, the minimization problem solving process of sparse coding was converted into a multi-recurrent neural network layer. As a result, sparse coding with a stronger representation capability was obtained. Then, overall end-to-end supervised training was implemented in combination with multi-scale spatial pyramid pooling (SPP). Additionally, GTSRB and affine transformation datasets were selected to verify recognition accuracy of the proposed approach. According to relevant experimental results, recognition accuracy and efficiency of the proposed approach are respectively up to 99.21% and 13fps at the GTSRB dataset. Moreover, objects deflected within an angle range of 60° can be effectively recognized.