In the recognition process of traffic signs by traditional convolutional neural networks, the discarding of the target coordinate frame by the maximum pooling layer makes it impossible to detect the instantiated parameters, such as the pose and the position of object features in real-world scenarios. To solve this problem, a CapsNet algorithm for the recognition of traffic signs based on the Bagging integration was proposed in this paper. Using CapsNet as the base classifier, a novel traffic sign recognition algorithm based on the Bagging integration framework was constructed. By setting an adaptive oversampling rate, the SMOTE algorithm was used to oversample the minority class samples to reduce the imbalance in the samples. Algorithms such as Multi-Scale CNN、Committee of CNN & MLP were selected under the same testing environment to compare and analyze the recognition accuracy of the German Traffic Sign Recognition Benchmark (GTSRB) and the multi-view dataset after the affine transformation. In the meantime, by using GTSRB as the training set and the sampled dataset of LISA and Belgium TS as the testing set, the generalization capability of the proposed algorithm was investigated. The experimental results showed that, compared to the optimum public mode, the accuracy of the traffic sign recognition model proposed in this paper was 99.07 % and the recognition accuracy on the multi-view traffic sign dataset was 77.58 %, which is equivalent to a 7.21 % enhancement in addition to a good generalization capability.

CICTP