Zero-shot image classification using coupled dictionary embedding

Zero-shot learning (ZSL) is a framework to classify images that belong unseen visual classes using their semantic descriptions about the classes. We develop new ZSL algorithm based on coupled dictionary learning. The core idea enforce features and attributes of an image share same sparse representation in intermediate embedding space, modeled as shared input space two sparsifying dictionaries. In training stage, we use from number seen for which have access both train dictionaries can represent feature vectors single vector. testing stage absence labeled data, are mapped into attribute by finding joint-sparse representations solely via solving LASSO problem. then classified given also provide attribute-aware transductive formulations tackle “domain-shift” “hubness” challenges ZSL, respectively. Experiments four primary datasets VGG19 GoogleNet features, provided. Our performances 91.0%, 48.4%, 89.3% SUN, CUB, AwA1 datasets, AwA2 57.0%,49.7%, 71.7%, respectively, when used. Comparison with existing methods demonstrates our method effective compares favorably against state-of-the-art. particular, leads decent performance all datasets.2