Zero-shot Hashing with orthogonal projection for image retrieval

Abstract

Hashing has been widely used in large-scale image retrieval. Supervised information such as semantic similarity and class label, and Convolutional Neural Network (CNN) has greatly improved the quality of hash codes and hash functions. However, due to the explosive growth of web data, existing hashing methods cannot well perform on emerging images of new classes. In this paper, we propose a novel hashing method based on orthogonal projection of both image and semantic attribute, which constrains the generated binary codes in orthogonal space should be orthogonal with each other when they belong to different classes, otherwise be same. This constraint guarantees that the generated hash codes from different categories have equal Hamming distance, which also makes the space more discriminative within limited code length. To improve the performance, we also extend our method with a deep model. Experiments of both our linear and deep model on three popular datasets show that our method can achieve competitive results, specially, the deep model can outperform all the listed state-of-the-art approaches.