Face recognition is an interesting and challenging area of image processing. Human brain can recognize faces almost instantly and existence of variations in face such as expressions, pose, head rotation or illumination affects this capability only to some extent. For a computer, however, these variations can dramatically change the identification of the person. Scientists have been researching on improvement of this area for more than 20 years but reliable facial recognition systems were not used in real world applications until recently. Moreover, face recognition requires huge amount of calculation. The process starts with detecting a face in an image, then reducing noise by cropping the picture to the face area, next extracting facial features, and finally compare the face to a set of provided face images called the training set. A computer with a lot of processing power is needed to do the whole calculation at a reasonable speed. With the recent advances in technology, embedded computers are getting smaller and more powerful; however, the question remains whether these small computers are powerful enough to be able to run the face recognition process in a real-time or near real-time fashion. This thesis provides an evaluation of three open-source widely used facial feature extraction methods, eigenfaces, fisherfaces and local binary patterns histograms, on three embedded hardware platforms, Raspberry Pi 2, Intel Next Unit of Computing (NUC) and AMD G-series system on chip. Moreover, it takes six different cases in which variations in face images are investigated into consideration. In order to have thorough results, this study performs all experiments twice, one time with two people as known ,i.e. their pictures exist in the training set, and one as unknown and another time with three people as known and two as unknown. Results are evaluated and discussed about for each case and each hardware platform separately, then an average of all experiments are reported as the overall performance for each method. Experiments show that eigenfaces had the highest performance among all, local binary patterns histograms had slightly lower accuracy of recognition and fisherfaces performed with the accuracy of 13,92 percent lower than eigenfaces.

Top