Zero-shot learning based on spatial pyramid matching using sparse coding
1. School of Information and Control Engineering, China University of Mining and Technology, Xuzhou, Jiangsu 221008, China; 2. Xuzhou Financial and Economic School, Jiangsu Union Technical Institute, Xuzhou, Jiangsu 221008, China
Abstract:The lowlevel visual features of images to train attribute classifiers are used in the attributebased zeroshot learning methods generally, and the corresponding classification accuracy heavily depends on specific lowlevel features. A zeroshot learning method was proposed based on spatial pyramid matching using sparse coding (SSPM_IAP). The structure flow chart of system was given, and the SIFT features were extracted from original images. The features were extracted and reduced by spatial pyramid matching using sparse coding and max pooling. The exacted image features were used to train the attribute prediction model. The SSPM_IAP algorithm was proposed to predict image attributes and classify images under the zeroshot setting. The comparison experiments were completed on Shoes and OSR datasets. The results show that compared with several popular zeroshot learning methods, the experimental time consumption is reduced, and the proposed method can achieve more accurate attribute prediction and better zeroshot image classification.