Face liveness detection based on IQA using anova feature selection

In the past few decades, there has been a growing interest in Facial Biometric systems that became a trend in a wide range of technologies like security, access control and surveillance applications. However, Spoof attacks remain the main challenge faced by facial biometric systems. A spoof attack arises when an individual attempt to disguise as someone else by a fake face to get an unauthorized access to the system, a fake face could be a photograph, dummy face or even a video display. To overcome these attacks on such systems, face liveness detection has been produced. There are various ways to detect the face liveness such by texture, motion analysis, determine a scenic clue or by using a thermal sensor. Two methods of detection were identified based on the necessity of user’s cooperation with the system. One is known as intrusive which requires user interaction with the system such in motion detection and the other is non-intrusive were no user effort is needed. For this purpose, image quality assessment has been utilized in the literature for face anti-spoofing detection. Image quality measures (IQMs) are efficient, user friendly, non-intrusive, low cost and present a low degree of complexity in implementation. However, they exhibit some limitations in terms of accuracy and efficiency of the system. Thus, an effective face liveness detection system based on image quality measures has been proposed in this thesis. The system was designed to conquer the limitations of accuracy in a trade off with high and cost ineffective feature extractor. System’s effectiveness was evaluated and benchmarked with other existing related work on CASIA face anti-spoofing database and the expandability of proposed work was further proven on NUAA imposter database. The feature set was selected based on IQMs discrimination power. Analysis of variance (ANOVA) was the statistical tool used to identify these IQMs. ANOVA was applied to find the p-value and F-score for each of the measures. A low p-value (high F score) for a test refers to an evidence to reject the null hypothesis. Then a feature selection strategy was further implemented to minimize the number of measures. The output measures have been employed as a feature extractor to design and develop the face liveness detection system. Image classification for real and fake samples was implemented by support vector machine (SVM). The system is restricted to 2D images. The test results and evaluations have been implemented by the statistical analysis testing and by liveness detection system in terms of accuracy, half total error rate (HTER) and system’s efficeincy. Results have consistently revealed that the proposed method outperforms other detection techniques over different types of spoofing attacks and mediums. The detection accuracy of the system was increased by nearly 13% while the computational load was decreased by approximately 50 % as compared to the state-of-art. The contribution of this work is to ensure the simplicity of detection system and improves its accuracy along with efficiency.

Text FK 2019 15 ir.pdf Download (698kB)

Actions (login required)

View Item