Combining Feature Extraction and Clustering for Better Face Recognition

Abstract

In this paper, we study the performance of face clustering approaches using different feature extraction techniques. This study will highlight best practices for handling faces of terrorists and criminals in an approach which we are working on to trace and red flag potential cases. Given as input images containing faces of people, face clustering divides them into K groups/clusters with each group containing images expected to represent almost the same person. Face clustering is very important, especially in forensic investigations where millions of images are available in crime scenes to be investigated. We study the performance of face clustering by first choosing different feature extraction techniques to capture information from faces. Feature extraction techniques are employed to check which face representation works better in describing faces as input to clustering algorithms. We also used Rank Order clustering algorithm which is known for its good accuracy when clustering face images along with other traditional clustering techniques. We evaluated the performance of feature extraction techniques and clustering algorithms using four datasets (JAFFE, AT&T, LFW, and YaleB); each imposing different challenges for face clustering with varying image environment and for datasets of different sizes. These datasets challenge clustering algorithms and feature extraction techniques in run time and clustering accuracy. Experimental results show the effectiveness of Rank Order clustering in terms of accuracy for small datasets while its run time performance degrades for larger datasets. K-means performed poorly on the LFW dataset. OpenFace performed the best in describing face images, especially on large datasets compared to other feature extraction techniques. The latter method reported high accuracy margin that is big and acceptable feature extraction time.