LightAt-CNN: A Super Efficient CNN model based on Attention Mechanism

This paper presents a highly efficient Convolutional Neural Network (CNN) architecture for traffic sign recognition that achieves state-of-the-art accuracy while dramatically reducing model complexity. Our approach combines a streamlined four-layer CNN architecture with a parameter-free Simple Attention Mechanism (SimAM) and strategic placement of regularization techniques. To address class imbalance in the German Traffic Sign Recognition Benchmark (GTSRB) dataset, we implement a novel loss function that combines Cross-Entropy Focal Loss with Class-Wise Label Smoothing. The proposed model achieves 99.21 percent accuracy on the GTSRB dataset while requiring only about 50,000 parameters, representing a significant reduction compared to existing lightweight models that typically use 120,000 parameters or more. This dramatic decrease in model size without compromising accuracy makes our approach particularly suitable for deployment in resource-constrained environments and real-time applications.