Abstract
This study compares deep learning methods for automatic road sign recognition, a critical task for ensuring safety in autonomous driving and smart city technologies. We evaluate the performance of six convolutional neural network architectures (LeNet5, AlexNet, VGG16, GoogLeNet, ResNet18, and MobileNetV3) and two object detectors (YOLOv8n and Faster R-CNN). A key feature of the methodology is the creation of a hybrid dataset combining a standard Kaggle dataset with unique images of Kazakhstan's road infrastructure. The use of localized data allowed us to account for regional specifics, including bilingual signs (e.g., "Toqta / Stop") and challenging lighting conditions. Experiments revealed that the GoogLeNet model demonstrates the highest accuracy (approximately 98%) for classification tasks, outperforming older or lightweight analogs. In object detection tasks, the YOLOv8n detector demonstrated the best balance between accuracy (Precision around 0.97) and speed, making it optimal for real-time systems, while Faster R-CNN was slower and produced more false positives. The paper pays special attention to the problem of small object recognition: the results confirm that adding local examples significantly improves the algorithms' robustness to cultural and linguistic variations in the road environment. The work demonstrates the potential of using adapted lightweight models for implementation in car on-board computers.