Browse Books

Reviewer: Jonathan P. E. Hodgson

The aim of this book is to describe how one might construct an ideal face recognition system that is completely automatic; unsupervised; noninvasive, in the sense that no special requirements for poses or lighting are required; and robust, so that it works in an unconstrained environment, and manages pose shifts, occlusion, and lighting changes. The author suggests several potential new applications for such a system, including home monitoring and intelligent cars, where the interface customization would be tied to recognizing the driver rather than to recognizing the key used by the driver. A substantial portion of the book is devoted to an overview of cognitive and biometric background information. The author maintains that the design of an intelligent face recognition system will benefit from an understanding of the human cognitive system. The second chapter is devoted to this material, as well as a review of existing face recognition technology. Based on the results of the Face Recognition Vendor Test (FRVT), machine-based face recognition outperforms humans, but this is only for variations in lighting, not pose or context variation. The main new direction that the author pursues is the use of video for face recognition. While this introduces difficulties as a result of motion and the presence of complex backgrounds, it has potential benefits from temporal information, abundance of data, and the fact that the video covers only a constrained area. In particular, it allows the system to address cases where there may be more than one face and partial (transitory) occlusion. The author's system uses FaceVACS, which is a proprietary system that was ranked first in the 2002 FRVT. It comes with a software development kit (SDK), so it can work with other software. Although the central eye detection algorithm of FaceVACS is only described generically in the FaceVACS documentation, the author was able to calibrate the system to balance between false acceptances and false rejections. By using known bounds for the ratio of face dimensions to inter-eye distance, the system can limit the search area for a face to a rectangle determined by these bounds once a pair of potential eyes is detected. For the recognition phase, three approaches are integrated: image face detection (from FaceVACS), a face tracker, and a temporal filter. These latter two are possible because of the system's use of video. A critical aspect of the author's proposed system is the management of the database of known faces. The face database is clustered in two ways, with each cluster of images of a single (recognized) face being made up of sub-clusters. This structure reduces the incidence of cluster overlap amongst the large clusters. The database must be able to grow rapidly at the start of the system, and then be updated to allow for such things as aging on the part of those to be recognized. The architecture of the system is based on a state machine, which is described and pictured in the book. The system is typically implemented on a Webcam with a USB interface to a PC. The algorithms are written in C++, with the DirectX SDK for video capture, OpenCV for image processing, and the aforementioned FreeVACS SDK. Currently, the system is able to process one-to-two frames per second on a 2.4GHz PC, requiring the human subjects to slow down their actions. A somewhat surprising finding is that the presence of a background does not degrade performance. The system was tested against pure image face detection on images showing pose differences, expression changes, scale changes, lighting changes, and fast motion. In all cases, the system outperformed image face detection alone. The book performs a valuable service by gathering together an overview of face recognition, including comprehensive bibliographies at the end of each chapter. The description of the author's system, while clear, is not at the level of detail that would allow for a straightforward reconstruction of the system or one like it; this is a pity because the results are encouraging and should be pursued. Online Computing Reviews Service