Wavelet-based Image Compression Using Human Visual System Models

Recent research in transform-based image compression has focused on the wavelet transform due to its superior performance over other transforms. Performance is often measured solely in terms of peak signal-to-noise ratio (PSNR) and compression algorithms are optimized for this quantitative metric. The performance in terms of subjective quality is typically not evaluated. Moreover, the sensitivities of the human visual system (HVS) are often not incorporated into compression schemes. This thesis develops new wavelet models of the HVS and illustrates their performance for various scalar wavelet and multiwavelet transforms. The performance is measured quantitatively (PSNR) and qualitatively using our new perceptual testing procedure. Our new HVS model is comprised of two components: CSF masking and asymmetric compression. CSF masking weights the wavelet coefficients according to the contrast sensitivity function (CSF)–a model of humans' sensitivity to spatial frequency. This mask gives the most perceptible information the highest priority in the quantizer. The second component of our HVS model is called asymmetric compression. It is well known that humans are more sensitive to luminance stimuli than they are to chrominance stimuli; asymmetric compression quantizes the chrominance spaces more severely than the luminance component. The results of extensive trials indicate that our HVS model improves both quantitative and qualitative performance. These trials included 14 observers, 4 grayscale images and 10 color images (both natural and synthetic). For grayscale images, although our HVS scheme lowers PSNR, it improves subjective quality. For color images, our HVS model improves both PSNR and subjective quality. A benchmark for our HVS method is the latest version of the international image compression standard–JPEG2000. In terms of subjective quality, our scheme is superior to JPEG2000 for all images; it also outperforms JPEG2000 by 1 to 3 dB in PSNR. We would like to thank the National Science Foundation for their support; this research was made possible through their sponsorship. iii Acknowledgments This work would not have been possible without the guidance and support I received from Dr. Amy Bell. When my research led me to a wall, she always guided me to the next open door. Her constructive criticism–always offered with a smile–fueled our research and polished this thesis into its final form. I offer my gratitude to Dr. Brian Woerner and Dr. Lynn Abbott for serving on my committee. My DSPCL cohorts, Lakshmi Iyer and Manish Manglani, were my source of truth. The task was never complete without their judgment.