Facial Image Classification Using Rotation, Illumination, Scale and Expression Invariant Dense Features and ENN

Face Recognition is immensely proliferating as a research area in the paradigm of Computer Vision as it provides an extensive choice of applications in surveillance and commercial domains. This paper throws light upon the comparison of various dense feature descriptors (Dense SURF, Dense SIFT, Dense ORB) with each other and also with their classical counterparts (SURF, SIFT, ORB) using a novel technique for recognition. This proposed technique uses Laplacian of Gaussian filter for enhancement of the image. It applies various dense and classical feature descriptors on the enhanced image and outputs a feature vector. In order to achieve high performance, this feature vector is given to Fisher vector since Fisher Vector is a feature patch-aggregation method. Finally, extended nearest neighbor Classifier is used for classification over the orthodox k-nearest classifier. Experiments were carried out on three diverse datasets—ORL, Faces94, and Grimace. On scrutinizing the results, Dense SIFT and Dense ORB were found to be preeminent as measured by various performance metrics. 98.44 on Grimace, 98.15 on Faces94.