A Lightweight Convolutional Neural Network for Traffic Sign Recognition: Comparative Analysis and Integration of Xai Techniques
Kaushik Biswas, Fariya Sultana Prity, Firoz Ahmed · 2025
Recent developments in automotive technologies and computer vision have significantly impacted Traffic Sign Recognition (TSR) sectors, particularly in the development of Driver Assistance Systems (DAS). Existing models tend to be computationally heavy and lack of transparency. This study addresses these limitations by proposing a lightweight, interpretable customized Convolutional Neural Network (CNN). To evaluate its performance, we compared it against 10 commonly used pre-trained models on the Belgium Traffic Sign Dataset (BTSD), the German Traffic Sign Recognition Benchmark (GTSRB), and the extended GTSRB Plus datasets. The proposed model outperformed other pre-trained models and obtained an accuracy of 98.12 % on BTSD, 99.83% on GTSRB, and 96.53% on GTSRB Plus. Furthermore, to enhance the reliability and transparency of the model, XAI mechanisms, specifically Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME) were incorporated. Ultimately, the findings demonstrate the advantages of the lightweight model over complex pre-trained ones, while also emphasizing the necessity of XAI in improving safety in DAS applications.