Image Compression Effects in Face Recognition Systems

With the growing number of face recognition applications in everyday life, image- and video-based recognition methods are becoming important research topic (Zhao et al., 2003). Effects of pose, illumination and expression are issues currently most studied in face recognition. So far, very little has been done to investigate the effects of compression on face recognition, even though the images are mainly stored and/or transported in a compressed format. Still-to-still image experimental setups are often researched, but only in uncompressed image formats. Still-to-video research (Zhou et al., 2003) mostly deals with issues of tracking and recognizing faces in a sense that still uncompressed images are used as a gallery and compressed video segments as probes. In this chapter we analyze the effects that standard image compression methods - JPEG (Wallace, 1991) and JPEG2000 (Skodras et al., 2001) - have on two well known subspace appearance-based face recognition algorithms: Principal Component Analysis - PCA (Turk & Pentland, 1991), Linear Discriminant Analysis - LDA (Belhumeur et al., 1996) and Independent Component Analysis - ICA (Bartlett et al., 2002). We use McNemar's hypothesis test (Beveridge et al., 2001 ; Delac et al., 2006) when comparing recognition accuracy in order to determine if the observed outcomes of the experiments are statistically important or a matter of chance. Following the idea of a reproducible research, a comprehensive description of our experimental setup is given, along with details on the choice of images used in the training and testing stage, exact preprocessing steps and recognition algorithms parameters setup. Image database chosen for the experiments is the grayscale portion of the FERET database (Phillips et al., 2000) and its accompanying protocol for face identification, including standard image gallery and probe sets. Image compression is performed using standard JPEG and JPEG2000 coder implementations and all experiments are done in pixel domain (i.e. the images are compressed to a certain number of bits per pixel and then uncompressed prior to use in recognition experiments). The recognition system's overall setup we test is twofold. In the first part, only probe images are compressed and training and gallery images are uncompressed (Delac et al., 2005). This setup mimics the expected first step in implementing compression in real-life face recognition applications: an image captured by a surveillance camera is probed to an existing high-quality gallery image. In the second part, a leap towards justifying fully compressed domain face recognition is taken by using compressed images in both training and testing stage (Delac, 2006). We will show that, contrary to common opinion, compression does not deteriorate performance but it even improves it slightly in some cases. We will also suggest some prospective lines of further research based on our findings.