XceptionSELite: Lightweight and Efficient Network for Traffic Sign Recognition

Intelligent traffic systems count on detecting traffic signs to boost driving efficiency and safety. The paper introduces XceptionSELite, a novel traffic sign classification method that uses an improved Xception architecture enriched with Squeeze-and-Excitation (SE) blocks. For real-time applications, our proposed model strives for high classification accuracy while preserving computation efficiency. We trained and validated the model using the GTSRB dataset, achieving 94% validation accuracy and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2 \%}$</tex> testing accuracy. The preliminary step in this process is to identify the Region of Interest, which helps focus on a particular region where traffic signs are likely to be present, reducing processing time. SE blocks adaptively adjust channel-wise feature responses, directing the network's attention to more relevant features, while depthwise separable convolutions drastically lower computation costs by separating traditional convolutions. The classification report highlights the model's balanced performance, including impressive precision, recall, and F1-scores across traffic sign categories. The training and validation curves demonstrate minimal overfitting and good generalization, suggesting that the depthwise and SE block modifications are beneficial to the network.