EfficientZSR: A Lightweight Generalized Zero-Shot Recognition Framework

Within the realm of deep learning, Zero-Shot Learning (ZSL) presents a significant challenge, particularly in resource-constrained environments where efficient model inference is crucial. To tackle this problem, we present EfficientZSR, a lightweight generalized zero-shot recognition framework that integrates the Vision Transformer (ViT) as the teacher model and EfficientNet as the student model. Through the application of knowledge distillation techniques, EfficientZSR effectively inherits the powerful feature extraction capabilities of ViT while maintaining a lightweight structure. We performed evaluations across three standard datasets. (AWA2, CUB, and SUN) and achieved outstanding results across all three, further validating the effectiveness and broad applicability of EfficientZSR.