A novel approach based on fully connected weighted bipartite graph for zero-shot learning problems

Abstract

Zero-shot learning (ZSL) is a learning paradigm that tries to develop a recognition model to recognize mutually exclusive training and testing classes. To recognize mutually exclusive classes, some kind of correlation between training and testing classes are required. This paper proposed an inductive solution of the ZSL problem in two stages: (1) a supervised multiclass classifier is trained on the training set and further asked to classify the testing images to its nearest training class. (2) A mapping function, which maps training class to testing class is used to obtain the final class for each testing image. The correlation between seen classes and unseen classes are obtained using the mapping function. We have proposed a graphical mapping function based on a fully connected bipartite graph for mapping between training and testing classes. Each edge of the bipartite graph is assigned a weight calculated by exploiting the semantic space. The proposed model is evaluated over the three well-known ZSL datasets: AWA2, CUB and aPY and obtained 66.59%, 48.95%, and 32.91% mean accuracy respectively. The obtained f1 score of the proposed method is 0.675, 0.565 and 0.492 on AWA2, CUB and aPY dataset respectively.