Optimized Hybrid 1CNN with Late Feature Fusion and Batch Normalized Fully Connected Layer for Traffic Sign Classification

Abstract

The objective of this study was to develop a digital traffic sign recognition system using artificial intelligence, particularly deep learning, to improve recognition performance. Novel hybrid CNN architecture was proposed by combining two fine-tuned models through feature fusion before the fully connected layer. To enhance training stability and convergence speed, batch normalization was applied to the fully connected layer. The proposed approach improved the model‟s ability to extract discriminative features and achieved higher accuracy compared with single-model methods. The model was trained on an augmented dataset of 11,000 Ethiopian traffic sign images representing 156 unique classes. Using transfer learning, the convolutional bases of MobileNet and VGG16 were fine-tuned, and their extracted features were fused before the classifier layer. The batch normalization layers reduced over-fitting by simplifying the computational load of the fully connected layer. Experimental results demonstrated that the hybrid model outperformed the individual models, achieving a testing accuracy of 99.18% compared with the baseline accuracies of VGG16 and MobileNet, which were 96.82% and 89.95%, respectively. The hybrid model successfully integrated the strengths of both architectures, resulting in faster training performance and improved classification accuracy.