Sliding RandAugment: Dynamic Dual-Window Data Augmentation for Efficient Traffic Sign Recognition

Data augmentation is known to improve model generalization and robustness by acting as a form of regularization. However, the computational burden of applying multiple complex transformations per image can significantly increase training time, especially on CPU-limited or embedded platforms. In this work, inspired by RandAugment, we propose Sliding RandAugment, a dual-window strategy that dynamically selects between "simple" and "complex" transformation pools via sliding windows, balancing diversity and efficiency. When integrated with the lightweight DeepThin network, our method reduces training time per epoch and improves accuracy. We further evaluate the method on additional architectures, including Wide-ResNet, MobileNetV2, and SqueezeNet, and observe consistent improvements over RandAugment. Evaluated on GTSRB, ETSD, and CIFAR-10 datasets, Sliding RandAugment is a practical and effective approach for traffic sign recognition on devices where computing resources are limited.