Embedded Zero-shot Learning Algorithm Based on Semantic Attributes

Abstract

Zero-shot learning is a technique capable of recognizing target categories even when labeled samples for these categories are completely absent. Traditional zero-shot learning methods based on embedding models usually have a low utilization rate of semantic attributes and exhibit a bias towards seen classes during testing. Addressing this, we propose an embedding-based ZSL method grounded on semantic attributes. This method uses a spatial attention mechanism during the construction of the semantic attribute embedding space, enabling the model to focus on more distinctive attribute features within the images. Consequently, it can utilize these distinctive features for similarity classification. Furthermore, a category calibration loss function is introduced to assign a greater weight to unseen classes and a lesser weight to seen classes, aiming to reduce the bias towards seen classes during testing. Extensive experiments were carried out on three mainstream ZSL benchmark datasets. Compared with some of the existing classical algorithms, our method demonstrated improved results.