Fusion of Face Recognition Classifiers under Adverse Conditions

Fusion of Face Recognition Classifiers under Adverse Conditions

Norman Poh (University of Surrey, UK), Chi Ho Chan (University of Surrey, UK) and Josef Kittler (University of Surrey, UK)
Copyright: © 2014 |Pages: 21
DOI: 10.4018/978-1-4666-5966-7.ch010
OnDemand PDF Download:
List Price: $37.50


A face acquired by recognition systems is invariably subject to environmental and sensing conditions, which may change over time. This may have a significant negative impact on the accuracy of recognition algorithms. In the past, these problems have been tackled by building in invariance to the various changes, by adaptation, and by multiple expert systems. More recently, the possibility of enhancing the pattern classification system robustness by using auxiliary information has been explored. In particular, by measuring the extent of degradation, the resulting sensory data quality information can be used to combat the effect of the degradation phenomena. This can be achieved by using the auxiliary quality information as features in the fusion stage of a multiple classifier system, which uses the discriminant function values from the first stage as inputs. Data quality can be measured directly from the sensory data. Different architectures are suggested in this chapter for decision making using quality information. Examples of these architectures are presented and their relative merits discussed. The problems and benefits associated with the use of auxiliary information in sensory data analysis are illustrated on the problem of personal identity verification used in biometrics.
Chapter Preview

1. Introduction

Recognising faces under adverse conditions is one of the most important challenges for practical face recognition system because unfavourable imaging conditions affect the quality of face image. In data analysis, this problem is called data drift. By drift we understand that the observe feature vectors are subject to a number of transformations which include factors affecting data acquisition as a special case. Depending on the nature of the data, these factors relate to changes in the camera characteristics (e.g. for visible, infrared, or thermal camera), the unconstrained environmental conditions (illumination, background noise, clutter, atmospheric conditions), and / or the behaviour of the imaged object during data acquisition process (e.g. pose, motion, deformation, biological evolution such as aging).

One of the possible solutions is to develop condition-invariant face descriptor for recognition, but it is impossible to be robust to all conditions. Another solution is to collect face images in all the possible conditions that may be envisaged during the system operation. This approach ensures that the image that serves as a basis for solving the classification problem remains representative in future operation. This is somewhat difficult to accomplish, as, at the outset, it is not always possible to predict all the possible types of changes in advance. An alternative solution is to apply a suitable normalisation procedure by evaluating the factor causing a data drift. Various normalization procedures proposed in the literature have been shown to be very effective in stabilizing the data(Poh, Kittler, & Bourlai, 2010).

Another effective approach to dealing with the data drift is to use multiple experts. In other words, a set of solutions to the same classification problem is used instead of the individually best solution. It is well known that, if these experts provide diverse opinions about the points to be classified, the classification accuracy of the solution is improved. It is less well known that multiple classifier systems also improve the robustness of the solution to adverse conditions.

In this chapter, we pursue this particular approach to the data drift problem. We show that the effectiveness of the multiple expert approach can be enhanced by making the use of information about the data quality. By data quality we mean an objective measure of the data departure from its nominal characteristic. As already indicated, data drift is caused by various factors which will be reflected in the properties of the sensory signals. One can view these signal changes as changes in signal quality. In the normalization approach discussed earlier we attempt to reverse the signal changes by the application of preprocessing algorithms that aim to stabilize the data to be classified. In the multiple classifier system approach the idea is to express signal changes in the form of quality measures. These quality measures can then be used as auxiliary features in the multiple classifier system fusion. As a result, the fused system decision is influenced not only by the expert opinions regarding the respective hypotheses, but also by measures of the signal quality. We formulate the problem of multiple expert decision-making which incorporates quality information. We then show that the use of this auxiliary information leads to further improvement of the system performance under data drift. This is illustrated on data relating to personal identity authentication using facial biometric.

The chapter is organized as follows. In the next section we develop a theoretical framework for the quality based fusion of multiple classifier systems. In the formulation adopted the quality information is used as additional features. Accordingly, the decision making in the fusion stage of the resulting multiple classifier system is realized in an augmented feature space. In Section 3, we demonstrate this approach on a two class problem of face verification, where the data drift is caused by illumination changes. We show that the use of multiple experts results in performance gains over the best performing expert. These gains are further enhanced by incorporating quality information in the fusion process. Conclusions are drawn in Section 4.

Complete Chapter List

Search this Book: