Auto Data Augmentation for Testing Set

Testing phase augmentation is a fast way to further improve the performance of image classification when CNN (Convolutional Neural Network) is already trained for hours. Limited attempts have been made to find the best augmentation strategy for testing set. We propose a reinforcement learning based augmentation strategy searching method for testing phase augmentation. With the augmentation strategy, we augment each testing image and integrate features of its augmented images into one feature. The reinforcement learning method searches the best parameters in the augmentation strategy which is formed as a matrix in this paper. Using the proposed method, we achieve competitive accuracies on image classification and face verification.