Application of Contrast Embedding with Feedback Regulation in GZSL

Generalized Zero-Shot Learning is a world-class challenge. in real life, many scenes have been used and have broad prospects for development. GZSL's purpose is to identify new classes by transferring semantic knowledge from seen classes to unseen classes. Although the generation model learned by the feature generation method can synthesize the missing visual features in unseen classes to reduce the data imbalance in GZSL. It is disadvantageous to the classification of GZSL due to the absence of the original image feature information area. In order to solve this problem, a method of integrating the generated model with the embedded model has been proposed recently to produce a hybrid GZS L framework. This hybrid framework maps real samples and synthetic samples into an embedded space for classification. On this basis, we propose to add a feedback regulation mechanism to feed back the hidden features before contrast embedding classification into the synthetic network of unseen classes. We named it CEFWGAN. We performed validation evaluations on three official base data sets (FLO CUB SUN) in which some of the evaluation criteria were superior to the current approach.