A Distinguishable Pseudo-Feature Synthesis Method for Generalized Zero-Shot Learning

Figure 1

Illustration of DPFS. (a) DPFS consists of an embedding network, an attribute projection module (APM), a pseudo-feature synthesis module (PFSM), a preclassifier and a classifier. In stage 1, the embedding network and the preclassifier are jointly pretrained to extract distinguishable features for seen classes. In stage 2, the network synthesizes distinguishable pseudo-features for unseen classes through APM and PFSM. Then, the features and the pseudo-features are fed into the classifier for GZSL tasks. (b) APM Details. APM builds sparse representations based on attributes. (c) PFSM Details. PFSM creates feature representations and synthesizes distinguishable pseudo-features with the selected features, the base vectors, and the unseen class attributes. The outliers of candidate pseudo-features are disposed of to get distinguishable pseudo-features.