Image data augmentation method based on maximum activation point guided erasure

Abstract

Deep neural networks usually contain tens to hundreds of millions of orders of learning parameters that provide the necessary representation to solve various visual tasks. But with the increase of the representational ability, the possibility of over-fitting also increase, which bring about poor generalization. In this paper, we propose MA (Maximum Activation point processing) algorithm, a new image data augmentation method which is designed to improve the generalization ability of the model and reduce the risk of overfitting. During the training process, the most discriminative part of the input image is searched for, and the model is driven to search for the supplementary information of the most important feature information by erasing the maximum attention image block. During this process, training images with different occlusion levels are generated as new inputs to the network and the model continues to be trained. The image erasure method based on the maximum activation point guidance only needs to modify the input image, which can effectively improve the robustness of the model to occluded image recognition, and can be integrated with various network structures. The effectiveness of our method is verified on the Cifar10, Cifar100 and Fashion-MNIST datasets.