Pairnorm based Graphical Convolution Network for zero-shot multi-label classification

Abstract

Zero-shot learning transfers the knowledge from the seen labels available during training to the unseen labels. In this paper, we propose a Pairnorm based Graphical Convolution Network for zero-shot multi-label classification (ML-ZSLPGCN). The proposed approach uses the label features obtained from the images during training and semantic embedding for the unseen labels. The ML-ZSLPGCN first creates the features corresponding to the images, and the label aware module creates the feature vector of the labels corresponding to the seen labels using the attention region embedding. A graphical convolution network takes the feature vector of seen labels during training and semantic word embedding for the unseen labels as input and learns the classifier. The proposed approach uses a pairnorm-based normalization scheme to tackle the over smoothing problem in the graphical convolution network. The experimental results on the NUSWIDE and MS-COCO datasets show that the proposed approach provides significant performance in terms of precision, recall, and F1 score in comparison to state-of-the-art approaches. © 2022 Elsevier Ltd