基于域特定特征的CLIP提示优化算法OA北大核心CSTPCD

When the testing data and training data follow different distributions,the neural network can undergo domain shift.The goal of domain generalization(DG)is to solve this problem by learning a general model that can handle unknown domains.Previous methods can extract domain-invariant features by means of data enhancement or feature space alignment,but new domain-specific features can be generated in the process of extraction,resulting in poor model generalization performance.On this basis,a simple and effective framework ERCLIP(extracting and removing domain-specific features for CLIP)is proposed to realize the application of large-scale pre-training model CLIP in DG.ERCLIP can realize precise semantic description of images by actively extracting domain specific features and incorporating them into text prompts.The experimental results on the public datasets OfficeHome,VLCS,and PACS show that ERCLIP can realize the best results among all algorithms,with an average accuracy of 83.4%on OfficeHome,83.5%on VLCS,and 96.5%on PACS.