A Comparative Study of Hierarchical Matching Algorithms for Face Recognition

The similarity measure based nearest neighbour classifier is commonly used in object recognition or retrieval systems. The result of a query in such a system is in general the images from the database whose descriptors are closest to the descriptor of the query image. An important issue in large scale recognition systems is the computational burden caused by measuring the similarity of the features derived from the query image with the templates of the database objects. Hierarchical Agglomerative Clustering (HAC) algorithms are considered as a solution to this problem. In these algorithms, a dendrogram is formed down to top in the training stage. In the recognition phase, the query image moves on the dendrogram from the highest level to lower levels in order to find the best matched object. One of the most affecting issues in these algorithms is the strategy used for building the dendrogram. In this paper, different techniques adopted for this purpose are studied and compared within the framework of a face recognition system. Our experimental results demonstrate that using an appropriate merging technique, the average recognition time reduces while the performance of the system is not highly degraded.