Enhancing Generalized Zero-Shot Learning through Local Attention

Generalized Zero-Shot Learning (GZSL) has demonstrated crucial value in practical applications, as it can classify samples even if their category data has not been trained on. However, most current GZSL methods utilize pre-trained models to extract image features. Since the distribution discrepancy between GZSL datasets and pre-training datasets, the performance of the extracted features is limited, particularly as the feature distributions of different classes tend to be similar, making it difficult to set them apart. To address this issue, we have designed a novel feature enhancement model, which consists of three core modules: a feature generation module, a local attention feature enhancement module (LAFM), and a unique discriminator module (UDM). LAFM focuses on fine-grained details by employing an image-blocking strategy and optimizing semantic cycle-consistency loss, ensuring that each region block contains critical classification semantic information. Additionally, while the semantic regressor generates new attribute features, UDM calibrates the generated and real features, further refining the accuracy of feature representation. Experiments on two benchmark datasets demonstrate that the proposed method surpasses existing advanced methods.