Deep fusion convolutional neural network for accurate and carbon efficient traffic sign recognition

Abstract Objective In intelligent transportation systems, accurate and efficient traffic sign recognition is essential for enhancing road safety and traffic management. This study proposes DF-CNN, a lightweight Deep Fusion Convolutional Neural Network designed to achieve high classification accuracy while minimizing computational complexity and energy consumption in alignment with Green AI principles. Methodology The proposed model is trained and evaluated on the GTSRB dataset comprising 43 traffic sign classes. The architecture follows a sequential convolutional design with hierarchical feature integration, enabling efficient feature extraction with reduced model complexity. The framework incorporates preprocessing, optimized training, and fivefold cross-validation to ensure robustness and reliability. Results Experimental results demonstrate that the proposed DF-CNN achieves a classification accuracy of 98.49% ± 0.21. The model maintains stable convergence behavior and achieves this performance with approximately 162 K trainable parameters and ~ 29 M MACs, indicating strong computational efficiency. Conclusion The proposed DF-CNN provides a balance between high accuracy and low computational cost, making it suitable for real-time traffic sign recognition and resource-constrained environments. The model aligns with Green AI principles by reducing energy consumption while maintaining strong predictive performance.