Training-Time Image Augmentation on MobileNetV3-Small: A Controlled CIFAR-10 Pilot

Training-time image augmentation may improve classification under image corruption, but stronger policies can add computational cost and may change clean accuracy. This study compares random crop and horizontal flip, RandAugment, and AugMix using a CIFAR-adapted MobileNetV3-Small trained from scratch on CIFAR-10. Each condition was trained for 50 epochs with two seeds. We measured top-1 accuracy on the clean CIFAR-10 test set and the original 15 CIFAR-10-C corruption types at five severities, along with training wall time, source-image throughput, and peak allocated GPU memory on an NVIDIA GeForce GTX 1650 with 4 GB VRAM. Mean clean accuracies were 86.39% for crop-and-flip, 86.18% for RandAugment, and 86.02% for AugMix; mean CIFAR-10-C accuracies were 67.04%, 71.00%, and 77.82%, respectively. Mean training times were 13.2, 19.1, and 121.4 minutes. Run times varied substantially between seeds, particularly for RandAugment and AugMix. With two seeds, one architecture, one dataset, and one local system, these results are exploratory and do not support broad generalization or a novelty claim. This is an exploratory report and has not undergone journal or conference peer review.