NeighborMix data augmentation for image recognition

NeighborMix data augmentation for image recognition

연구 분야: Artificial Intelligence

학회: Multimedia Tools and Applications

Data augmentation can effectively enrich the diversity of training datasets to improve the generalization ability of deep learning models. Existing augmentation methods have achieved excellent performance on image recognition. However, the images generated by them can not well improve the accuracy and robustness of the model trained with natural scene data. In this paper, an augmentation method called NeighborMix is proposed to address this problem. NeighborMix only performs simple operations on the original image. Specifically, NeighborMix selects and pastes an occlusion object closely related to the original input image in a successive two-stage fashion. In the first stage, a block from a single original image is selected as the occlusion object. In the second stage, the occlusion object is copied and pasted into the original image to generate a new image. Finally, extensive experiments on benchmark datasets demonstrate that NeighborMix achieves top-1 results of 93.54%, 70.62%, and 89.30% superior model generalization performance on CIFAR-10, CIFAR-100, and STL-10, respectively, and performs the most robustness with a standard error of 0.05%, outperforms state-of-the-art data augmentation techniques.