Large-Scale Face Image Retrieval: A Wyner-Ziv Coding Approach

Great progress in face recognition technology has been made recently. Since the first face recognition vendor test (FRVT)Phillips et al. (2007) in 1993, face recognition performance has been improved by two orders of magnitude in thirteen years. Notably, in the FRVT 2006 it is the first time that algorithms are capable of human performance levels, and at false acceptance rates in the range of 0.05, machines can outperform humans Phillips et al. (2007). The advances are promising for face verification applications where a typical one-to-one match is performed. It is still a grand challenge to power large-scale face image retrieval. Large-scale face image retrieval is the enabling technology behind the next generation search engines (search beyondwords), bywhich web users can do social searchwith personal photos. High performance face identification algorithms are needed to support large scale face image retrieval. Compared with face verification, face identification is believed N times harder than face verification due to its nature of 1:N problems. The number of individuals N in the database has a great impact on both the effectiveness and efficiency of face identification algorithms. With the state of the art face identification algorithms, the identification rate is only around 70% (rank = 1) for the FERET database, a gallery of ten thousands individuals. When to serve for large-scale face image retrieval applications, the identification rate will further decrease as the gallery size increase (fortunately not linearly but logarithmically).The computing complexity of face identification is linearly related to the number of individuals N. For largescale face image retrieval the efficiency of face identification is a key issue. In this paper we focus on the efficiency aspects of face identification. Technically, it is very challenging to find a person from a very large or extremely large database which might hold face images of millions or hundred millions people. A highly efficient image retrieval technology is needed. Indexing technology based on tree structures has been widely used in commercial search engines. These structures are quite efficient for small dimensions (of the order of 1-10). However, as the data dimensionality increases, the query performance of these structures degrades rapidly. For instance, White and Jain report that as the dimensionality increases from 5 to 10, the performance of a nearest-neighbor query 2