Traffic Sign Recognition Using Lightweight Subnetworks of Deep CNNs with Transfer Learning

Traffic sign recognition plays a vital role in intelligent transportation systems by facilitating automated vehicle navigation and enhancing road safety. This study investigates the effectiveness of subnetworks of several advanced deep learning models—namely VGG16, ResNet50, EfficientNetB0, and a custom-designed convolutional neural network (CNN) on the German Traffic Sign Recognition Benchmark (GTSRB) dataset. To enhance model performance, techniques such as transfer learning, fine-tuning, and hyperparameter optimization were used. Initial training on the original dataset yielded high accuracy rates of 98.11% for the custom CNN, 98.92% for VGG16 subnetwork, 97.24% for ResNet50 subnetwork, and 97.97% for EfficientNetB0 subnetwork. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied, leading to improved accuracy scores of 99.06% for VGG16 subnetwork, 97.28% for ResNet50 subnet-work, and 98.03% for EfficientNetB0 subnetwork. VGG16 subnetwork consistently achieved the highest performance across both original and augmented datasets, highlighting its robustness in traffic sign recognition tasks. The results highlight the value of leveraging transfer learning, model fine-tuning, and data balancing strategies to optimize performance in deep learning-based recognition systems. This research contributes to the advancement of practical, reliable traffic sign recognition technologies in real-world scenarios.