Image classification robustness improvement method based on metric learning

• Main Authors: WANG XUYAO, ZHANG LIJUN
• Format: Patent
• Language: Chinese; English
• Published: 08.07.2022
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

The invention discloses an image classification robustness improving method based on metric learning. For a deep learning image classification task, an original data set corresponding to the image classification task is acquired, an adversarial sample data set is generated for the acquired original data set by a projection gradient descent method, the original data set and the generated adversarial sample are merged into a new data set, and then the new data set is input into a classifier for model training. And training until the model converges and then outputting. Wherein in the classifier model training stage, a sample distribution constraint module is added to an original classifier, and the sample distribution constraint module updates parameters of the classifier by applying a total loss function of metric learning. According to the method, distribution constraints and spatial features of normal samples and adversarial samples are considered, and the problem that the relation between sample distributio