Hidden assumption of face recognition evaluation under different quality conditions

Automatic face recognition remains a challenging task due to factors such as variations in recording condition, pose, and age. Many schemes have emerged to enhance the performance of face recognition to deal with poor quality facial images. It has been shown that reporting average accuracy, to cover a wide range of image quality, does not reflect the system's for any specific quality levels. This raises the need to evaluate biometric system's performance at each quality level separately. Challenging face databases have been recorded with varied face image qualities. Unfortunately, the performance of face recognition schemes under different quality conditions, reported in the literature, are evaluated under hidden assumption which cannot be achieved in real-life applications. In fact, this problem could be a source of attack that interferes with the verification through manipulating the recording condition. In order to remedy this problem, two requirements are to be imposed: 1) the matching criteria should be based an Adaptive Quality-Based Threshold (AQBT) and 2) at the verification stage the quality level of an input face image should be determined and classified into one of a non-overlapping predefined quality levels. We illustrate our idea by experiments conducted on the extended Yale B face benchmark dataset. Our experimental results indicate that if AQBT is not adopted, false rejection rates becomes very high (always reject) when using low quality face images.