Attention based Zero-Shot Learning for Unseen Class Classification

In the realm of computer vision and deep learning, recent strides have been made towards more robust image classification systems. However, traditional deep learning models encounter significant hurdles, including the demand for a large number of labeled data per category, the intricacy of distinguishing subtle differences between visually similar classes, and the failure to classify new classes. This study introduces a transformative approach to address these challenges: Zero-Shot Learning (ZSL). ZSL revolutionizes deep learning by training models to classify classes not present during training, employing attributes or semantic descriptions. This method seamlessly transfers knowledge from known to unknown categories by leveraging semantic cues or auxiliary data, eliminating the need for exhaustive labeled examples for every class used in analyzing the AWA2 and CUB datasets, standard benchmarks in the ZSL domain. The performance of the proposed model is analyzed using harmonic mean. The proposed model demonstrated significant improvements in generalization to unseen classes.