Effects of lossy compression on face recognition algorithms

Abstract

Face databases can consist of a few hundreds of face images to thousands, even millions. Because of storage and banwidth limitations, face databases are maintained under compressed domain. One of the related problems is the performance evaluation of traditional face recognition techniques on the compressed face images. The effects of information loss due to the compression, on the performance of principal face recognition techniques, the most robust face recognition technique against compression, the extend to which face images can be compressed without a major performance deterioration and the most appropriate compression technique for face images are determined. It is concluded that the face images can be compressed to 100:1 with face-specific compression techniques, 40:1 with SPIHT technique and 20:1 with VQ, JPEG and JPEG-2000 techniques. Most robust face recognition technique against compression is "Fisherface" method. The eigenfaces generated from compressed face images at 0.4 bit/pixel rate performed better recognition than eigenfaces generated from non-compressed images for VQ, JPEG and JPEG-2000 techniques.