Semantic Attribute Mapping for Zero-Shot Object Classification in Deep Learning Networks

One of the successful approaches to deal with the issue in object classification is the ability of the model to recognize categories that it has not seen before, referred to as the zero-shot learning or ZSL. This work aims at training and testing ZSL methods on a standard benchmark known as Animals with Attributes 2 (AwA2). In this work, CNN, to be precise ResNet is used for feature extraction after which a semantic attribute of each class is used for classification. These then are the semantic attributes of objects The ZSL model aligns the visual features from unseen objects to these attributes for classification. Evaluations based on top 1 and the top-5 accuracy to know if the model is likely to generalize to other classes. To ensure the system is computationally efficient and scalable, TensorFlow/Keras is used for the CNN features, and Hugging Face transformers for attribute features.