Supervised two-step hash learning for efficient image retrieval

Content-based image retrieval (CBIR) attracts more and more interests in modern applications. Hashing method is a popular solution of CBIR. Among all the hashing methods, supervised deep learning approaches have received brilliant advantages encouraged by the rapid development of convolutional neural networks in recent years. In this paper, we propose a supervised two-step hash learning method that demonstrates high accuracy and fast speed. Our technical contributions include a feature preparation part and a two-step hash learning process with a carefully designed prototype code system for utilizing supervised labels. Our method achieves satisfactory results via a quite short training time. We can extract well similarity-preserving features, learn a comprehensive function mapping and get compact hash codes as well. Experiments are conducted on some widely-used public benchmarks MNIST and CIFAR-10, indicating that our proposed method outperforms several state-of-The-Art methods by significant improvement.