Face Attributes Retrieval by Multi-Label Contractive Hashing

Substantial increase of Internet data requires efficient storage and rapid retrieval strategy. Hence, supervised hashing method is introduced in this issue. By mapping high dimensional data to compact binary codes, supervised hashing methods could downsize data while preserving semantic similarity based on labels. However, most of these hashing methods are designed for simple binary similarity, therefore they fail to manage the complex multi-level semantic structure of multi-label images. In this work, we propose a novel Multi-Label Contractive Hashing (MLCH) to preserve multi-level semantic similarity of face attributes images. To improve the efficiency of training process, an optimized triplet selection algorithm is implemented. Gradual learning is adopted to accelerate the rate of convergence and enhance the performance of proposed model. Meanwhile, contractive constraint is introduced to obtain more saturated binary codes. The proposed MLCH is evaluated on datasets CelebA and PubFig. Experimental results prove the validity of these ingenious strategies and demonstrate superiority of MLCH to the state-of-the-art hashing methods in large-scale image retrieval.