Label Projection based on Hadamard Codes for Online Hashing

When new data streaming arrives, traditional hashing methods should retrain the functions based on all samples. That leads to high training time complexity. In contrast, online algorithm re-computes just arrival and has been widely applied in large-scale image retrieval tasks. However, differences exist numbers labels between old datasets, which causes imbalance problem while establishing their similarity matrix. This paper proposes a novel supervised method, Label Projection, Hadamard Codes for Online Hashing (LHOH), jointly employs label projection preservation mechanism solve problem. addition, LHOH considers codes as target domains avoid of difficult discrete optimization objective function. Then, matrix weight values, can computing datasets preserve consistency Hamming semantic space similarity. To increase distinguishability among hash codes, designs triple supervision learning mechanisms, including assigning projecting labels, embedding labels. validate performance proposed this sets up approximate nearest neighbor (ANN) search comparative experiments two used datasets. The final results show that outperforms six current state-of-the-art methods.