Semi-supervised model generalization method fusing label denoising

• Main Authors: XIE HAN, LIU WENXI, YU YUANLONG, LIN LUOJUN
• Format: Patent
• Language: Chinese; English
• Published: 14.04.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; IMAGE DATA PROCESSING OR GENERATION, IN GENERAL; PHYSICS

The invention relates to a semi-supervised model generalization method fusing label denoising. The method comprises the following steps: step 1, generating a pseudo label of label-free source domain data by using a pseudo label generation model; 2, a dual calibration generalization model is adopted to learn on source domains of pseudo labels and real labels, clean samples are selected for exchange according to a small-loss strategy, the two parties update the screened clean samples, and meanwhile, a style confusion module is inserted to improve the generalization ability of the model; and step 3, the intermediate domain comprises a labeled source domain and a label-free source domain with clean samples, and the progressive intermediate domain generation module linearly mixes the two samples and enters the next cycle to serve as new labeled source domain data of a pseudo-label generation model. The method is beneficial to improving the generalization ability of the model. 本发明涉及一种融合标签去噪的半监督模型泛化方法，包括：步骤1、利用伪标签生成