Attention based Zero Shot Learning

Abstract

The purpose of zero-shot learning (ZSL) is to build a classification model which beneficial for no labeled samples of classes. The goal of ZSL is to recognize unseen objects in the training. For this purpose, seen and unseen classes required semantic attributes. After providing the auxiliary semantic attributes, ZSL predicts the unseen object by searching its label for a similar semantic attribute class. For this purpose, Latent Feature Guided Attribute Attention (LFGAA) framework have been proposed which is related to category features and visual semantic information to remove the ambiguity issues. In this framework, feature space and embedded subnet learn the projections at the same time. First, I select images from datasets i.e., CUB, SUN, and AWA2. After that, apply pre trained network ResNet-101 as a backbone. Then, I generate different regions using the attention layer and attention masks. In the end, embedded Subnet is generated from different regions using LGA. To perform the object-based attribute attention, Latent Guided Attention (LGA) is used in the middle layers of an embedded subnet. In Inductive and Transductive ZSL setting, experiments are performed on AWA2, CUB and SUN datasets using Latent Feature Guided Attribute Attention (LFGAA) framework with the help of Anaconda platform. In case of Inductive, the achieved accuracies are 74.227, 79.220 and 59.167 using AWA2, CUB and SUN datasets respectively. In case of Transductive, the achieved accuracies are 82.602, 82.476 and 58.056 using AWA2, CUB and SUN datasets respectively. Our proposed Latent Feature Guided Attribute Attention (LFGAA) method outperform state of the art methods.