Title: Graph-based One-step Hashing for Image Nearest Neighbor
Abstract: In large scale image retrieval domain, hashing method is becoming more and more significant for fast approximate similarity search and data storage efficiency. One problem of the previous hashing method is that most current methods are two-step strategy hashing methods, i.e., quantization and then binarization, which may achieve sub-optimal results because even each step achieve its optimal, the final result maybe suboptimal. Another concern is that most methods only preserve global or local similarity structure to reduce the loss of quantization. To solve the above problems, in this paper, an effective one-step hashing method is designed to achieve the quantization and the binarization result in one-step and preserve the global and local similarity structure to enhance hashing performance. Specially, a linear transformation matrix is used to transfer original image into binary codes which will guide the transformation matrix in return. Meanwhile, global and local similarity are imposed into the guidance of transformation matrix learning as well.