Implementation of Multi-Centroid Moment Invariants in Thermal-Based Face Identification System

Abstract

Problem statement: Current paramount methods and approaches in face identification field rely on facial characteristics; such as location of eyes, length of nose and mouth, regardless the type of medium used to acquire the facial images. The visibility of these facial characteristics varies significantly with environmental factors (e.g., lighting elements). Various researches have been devoted to develop methodologies for addressing these problems. Despite the overwhelming effort to overcome these problems, the current framework of face identification research has a conceivable weakness due to its very nature. Approach: In this study, we presented a novel framework for face identification based on thermal information extracted from facial images acquired within the thermal spectrum. The motivation initiating to conduct this research was to engage in extracting significant facial characteristics from the acquired thermal information, which differs from current facial characteristics that are visible over the skin. Primarily, we proposed a non-holistic analysis approach by decomposing the input image into several input images via multiple threshold method (3 threshold values). This allowed us to analyze the thermal information given within a specific thermal range. Then, the first moment invariant, I1, from Hu’s classical moment invariants were computed and extracted (with respect to centroid point obtained from frontal shot source images) which constituted the feature database. For each subject, we used one (frontal shot) registered image with two and four test images (left, mid left, right and mid right). This approach employs minimum distance measurement method for classification purposes. Results: We had conducted experiments on the IRIS IR facial dataset with encouraging results of 95% of correct identification rate for test to registered image ratio of 2:1 and 92.5% of correct identification rate for test to registered image ratio of 4:1. Conclusion: The obtained results demonstrated the high capability of Hu’s classical moment invariants as a feature in thermal based face identification and introduce new ways for classical methods to further be utilized in theoretical and empirical research in the area.