Color-Spatial Autoaugment another Approach to Autoaugment Policies Found & Implementation

Abstract

Data augmentation is a highly effective technique for improving modern image classifiers’ accuracy and has continuously improved over the years. In this paper, we propose a different approach to implementing AutoAugment policies known as ”Color-Spatial AutoAugment,” Our implementation utilizes the best policies dis- covered by AutoAugment for specific datasets. It categorizes them into color and spatial, thereby improving image classification accuracy. Applied to CIFAR-10, CIFAR-100, and SVHN datasets, our method significantly improved CIFAR-10 and CIFAR-100, achieving top-1 accuracies of 91.1% and 60%, respectively. These results mark a 5.54% improvement on CIFAR-10 and a substantial 7.05% increase on CIFAR-100 over the traditional AutoAugment. On the SVHN dataset, how- ever, our approach was a bit short, yielding a top-1 accuracy of 91.15% compared to AutoAugment’s 93.55%. These findings highlight the potential of Color-Spatial AutoAugment and the improvement over AutoAugment within the same training conditions.