Siamese-twin random projection neural network with Bagging Trees tuning for unsupervised binary image hashing

Abstract

In this paper a Siamese-Twin Random Projection Neural Network (ST-RPNN) is proposed for unsupervised binary hashing of images. ST-RPNN is made of two identical random projection neural networks with hard threshold neurons where the binary code is taken as the neuron outputs. The learning objective is to produce similar binary codes for similar input image pairs and different binary codes otherwise. The learning process is divided into two steps. Firstly, overcomplete random projection is used to produce a sufficiently long code, followed by a fast sparse technique for neurons selection (FSNS). Bootstrap Aggregation Trees or Bagging Trees (BT) is then used to make a refined compact code section. BT is also used as a fast retrieval tool that ranks the database with respect to a query without distance calculations and with a significantly lower complexity than Hamming distance approach. The proposed technique is compared with 10 unsupervised image binary hashing techniques on the COREL1K dataset and the CIFAR10 dataset. The proposed technique obtained better precision-recall results than all compared techniques on the COREL1K dataset, and better than 8 of them on the CIFAR10 dataset.