Abstract:In order to alleviate the domain shift problem in traditional zero-shot image classification models, a zero-shot image classification model based on unknown-class semantic constraint autoencoder is proposed. Firstly, a pre-trained ResNet101 network is used to the extract visual features of all known-class and unknown-class images. Then, the extracted deep visual features are mapped from the visual space to semantic space through an encoder. Next, the obtained semantic vectors after mapping are reconstructed into the visual feature vectors through a decoder. In the training process of semantic autoencoder, a constraint is imposed based on the distribution alignment between the clustering visual centers of unknown-class images and the unknown-class semantic prototypes to alleviate the domain shift problem. Finally, based on the similarity between the semantic vector of testing image predicted by encoder and the semantic prototype of each testing class, the nearest neighbor algorithm is used to achieve zero-shot image classification. Experimental results on AwA2 and CUB datasets show that the proposed model has high classification accuracy.