Abstract: This paper addresses the problem of approximate nearest neighbors (ANN) search in large-scale image collections. Inspired by the idea of multi-view observation in daily life, we propose a novel unsupervised hashing method to solve large-scale image retrieval on the scenarios of single information source, dubbed Multi-view Ensemble Hashing (MEH). MEH is realized by ensemble learning and a parallel architecture. In our approach, MEH learns a set of convolution filters from abundant images by principal component analysis (PCA) off-line at first. Next, MEH filters the original image collection of single information source respectively via these convolution filters, to generate the multi-view data itself. Then, MEH uses a traditional hashing method to learn hash function and hash code respectively in each generated view. Finally, MEH merges the results of multi-view together to achieve a final retrieval result by voting. Extensive experiments on dataset CIFAR-10 and LabelMe show the superiority of our proposed approach over several state-of-the-art hashing methods. Compared to the original hashing methods that used as the operator in MEH, our proposed approach improves the retrieval precision over 100% at code size of 16-bit, and 10% at code size of 256-bit. Furthermore, the cost of MEH maintains an approximate level for its parallelizable structure.