Design of an Augmentation Model for Facadeskull Faceoverlay Using Deep Key-Point Matching

Abstract

Recognition offacial data points from images has multiple applications, which include, but are not limited to, geometry analysis, mood analysis, drowsiness analysis, skull disfigurement analysis, etc. In order to perform this task, a wide variety of image processing operations are needed, which include, but are not limited to, face localization, cascade object detection, facial geometry analysis, etc. In this text, a novel method for mapping facial key-points with computer tomography (CT) images is proposed, which assists in estimation of skeletal disfigurement. Over the years, a large number of approaches are proposed by researchers, but they showcase limited performance in terms of accuracy of mapping, and accuracy of inter-image alignment. In order to remove this drawback, this text proposes design of a novel machine learning model that combines facial key-points with relevant CT key-points for face-to-CT mapping. A custom layered convolutional neural network (CNN) architecture is used for mapping facial data with its relevant key-points, and the same architecture is extended for mapping CT imagery with its key-points. Both the key-point pairs are then mapped using a linear classification engine, wherein correlation between key-point locations is used for skull-face overlay. This correlation assists in finding the best key-point pairs, which allows for merging facial images to their CT image counterparts. The current architecture is capable of merging any frontal facial image with its frontal CT scan with 91% accuracy, and requires a very low delay. Due to such a high performance, the proposed model outperforms most of the recently proposed methods, thereby making it useable for clinical usage. The proposed model was tested for a wide variety of image variations, including illumination, sizing, skin colour, etc. and a consistent performance was observed, thereby suggesting good scalability.