Hybrid Image Improving and CNN (HIICNN) Stacking Ensemble Method for Traffic Sign Recognition

Traffic sign recognition techniques aim to reduce the probability of traffic accidents by increasing road and vehicle safety. These systems play an essential role in the development of autonomous vehicles. Autonomous driving is a popular field that is seeing more and more growth. In this study, a new high-performance and robust deep convolutional neural network model is proposed for traffic sign recognition. The stacking ensemble model is presented by combining the trained models by applying improvement methods on the input images. For this, first of all, by performing preprocessing on the data set, more accurate recognition was achieved by preventing adverse weather conditions and shooting errors. In addition, data augmentation was applied to increase the images in the data set due to the uneven distribution of the number of images belonging to the classes. During the model training, the learning rate was adjusted to prevent overfitting. Then, a new stacking ensemble model was created by combining the models trained with the input images that were subjected to different preprocessing. This ensemble model obtained 99.75% test accuracy on the German Traffic Sign Recognition Benchmark (GTSRB) dataset. When compared with other studies in this field in the literature, it is seen that recognition is performed with higher accuracy than these studies. Additively different approaches have been applied for model evaluation. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to make the model explainable. Evidential deep learning approach was applied to measure the uncertainty in classification. Results for safe monitoring are also shared with SafeML-II, which is based on measuring statistical distances. In addition to these, the migration test is applied with BTSC (Belgium Traffic Sign Classification) dataset to test the robustness of the model. With the transfer learning method of the models trained with GTSRB, the parameter weights in the feature extraction stage are preserved, and the training is carried out for the classification stage. Accordingly, with the stacking ensemble model obtained by combining the models trained with transfer learning, a high accuracy of 99.33% is achieved on the BTSC dataset. While the number of parameters the single model is 7.15 M, the number of parameters of the stacking ensemble model with additional layers is 14.34 M. However, the parameters of the models trained on a single preprocessed dataset were not trained, and transfer learning was performed. Thus, the number of trainable parameters in the ensemble model is only 39,643.