Fourier Series Model for Facial Feature Point Land-Marking

Abstract

The field of digital health apps, combined with intelligent learning systems, is new and expanding to incorporate a wide range of possibilities in different domains. An application in the field of digital therapy is for the incorporation of emotion recognition systems as a tool for therapeutic interventions. Adopting an individually tailored virtual world combined with a novel reward system in a gaming scenario, complemented with the technical affinity of most autism spectrum disorder (ASD) children makes a suitable atmosphere for therapeutic intervention. In this paper the use of image processing techniques coupled with Fourier models is used to generate point land-mark annotations on facial features in an image. The OULU-CASIA database was used for the analysis process. The images were first pre-processed based on previous work to reduce background noise and focus on the face. Afterwards a de-correlation stretch was executed to separate different features. A series of morphological, region detections and boundary traces followed. Fourier series models were used to transition the rough segmented pixel data into a smooth geometric representation. Twenty evenly distributed land-mark points are then selected from a fine mesh. Results showed that the geometric representation adhered to the segmented pixel data with a mean of 81.88% Dice similarity. The positive outlook highlighted the effectiveness of such a technique in automating the land-mark annotation process, which is tedious and time consuming. This method leads to explainable machine learning feature representations, which lead to more robust emotion recognition models.