Simultaneous feature learning and hash coding with deep neural networks

摘要: Similarity-preserving hashing is a widely-used method for nearest neighbour search in large-scale image retrieval tasks. For most existing methods, an first encoded as vector of hand-engineering visual features, followed by another separate projection or quantization step that generates binary codes. However, such feature vectors may not be optimally compatible with the coding process, thus producing sub-optimal In this paper, we propose deep architecture supervised hashing, which images are mapped into codes via carefully designed neural networks. The pipeline proposed consists three building blocks: 1) sub-network stack convolution layers to produce effective intermediate features; 2) divide-and-encode module divide features multiple branches, each one hash bit; and 3) triplet ranking loss characterize more similar second than third one. Extensive evaluations on several benchmark datasets show simultaneous learning brings substantial improvements over other state-of-the-art unsupervised methods.