Distance Metric Learning for Semantic Segmentation based Graph Hashing

Similarity search in large-scale dataset has been the key technology in the fields of computer vision and multimedia in recent years. In order to support fine-grained image analysis and image retrieval, semantic segmentation based multi-instances search is becoming an urgent research issue. In this paper, a distance metric learning for semantic segmentation based graph hashing (MLSS) approach is proposed, which learns an optimal distance metric by minimizing the objective function to preserve the semantic similarity between the labels and instances. Then, an alternating optimization strategy is designed for the objective function by using gradient descent method and Rayleigh-Ritz theorem. Finally, we use the anchor graph hashing method to hash all instances into binary codes for similarity search. Multiple experiments are conducted on the public datasets and the comparative analysis validate the advantage of our proposed approach over the existing hashing methods.