An Improved Hashing Method for Image Retrieval Based on Deep Neural Networks

Hashing algorithm projects the vector of features onto the binary space that generate the binary codes to reduce calculating time. Thus Hashing Algorithm is widely used to improve retrieval efficiency in traditional image retrieval methods based on Deep neural networks (DNNs). In this paper, we extract the feature vectors whose elements between 0 and 1 by DNNs and linear scaling method, then we define the mean of each column vector of the matrix consisted of these feature vectors as threshold to create corresponding hashing codes after two-stages binarization. Since threshold brings major effect to the preservation of the similarity between images, during this process, the two-stages binarization play two important roles: 1) optimizing thresholds; 2) optimizing hash codes. The promising experimental results on public available Cifar-10 database show that the proposed approach achieve higher precision compared with the state-of-the-art hashing algorithms.