Semi-hard constraint augmentation of triplet learning to improve image corruption classification

When facing the challenge of image distribution shift and natural corruptions, most of data augmentation methods only consider the diversity of training image to enlarge the data quantity, the hardness quality of augmented training image has not been explored. In this paper, we propose the semi-hard constraint augmentation (ShCA) of triplet learning method to improve the image corruption classification. First, the semi-hard positive and negative samples are adaptively generated with the online instance distance ranking. The hardness of augmented image is manipulated by the amplitude disturbance term of selected information images, this forms the relative narrower margin distance between positive and negative images without model collapse. Second, the semi-hard triplet learning can promote the subtle discriminative ability of reserved phase features in the smaller optimization space, which fine structure features can make the model corruption robustness. Compared with other state-of-art methods on the four public clean and corrupted image classification datasets, our proposed ShCA method gets the better performance of image corruption classification, especially for the corrupted images with noise, blur, weather and digital categories. The codes are available at https://github.com/zhangxhxh/ShCA.