Multi balanced trees for face retrieval from image database
Pengyi Hao, Sei‐ichiro Kamata · 2011
We are interested here in retrieving images containing a specific person in image database. Due to large variations in illumination conditions, hairstyles, facial expressions, etc. and the factors like occlusion, sunglasses, profile, etc., robust face matching has been a challenging problem. On the other hand, the speed of search is also a considerable issue, especially for the dataset with millions of face images. Inspired by face tracks in video retrieval which take advantages from the abundance of frames to get multiple exemplars, we present an approach named multi balanced trees for face retrieval from image dataset in this paper. Face images in the dataset are efficiently organized by the trees produced for persons. Multi sampling on the facial components employs the rich local information, which can help to differentiate different persons. Given a query face, a sorted face set with similarities is obtained by inserting the query into a tree. It is easy and fast to get the search results in respect that it avoids calculating the distances between query and elements in the cluster. In addition, a rectification strategy is given in the query process to rectify the error occurred in the generation of trees, resulting in a significant improvement of retrieval quality. Experimental results show the better face grouping ability in comparison with traditional methods. The speed of searching is improved as well.