Lossy compression of head and shoulder images using zerotrees of wavelet coefficients
D.B. Gerhard, Witold Kinsner · 2002
This pager presents an image compression system based on the wavelet transform. The compression system includes preprocessing to remove unimportant information, wavelet transform of the image, thresholding of the wavelet coefficients, zerotree coding of the thresholded coefficients, and finally arithmetic coding of the zerotree code. Post-processing is done to remove salt and pepper noise, an artifact of the wavelet transform at high compression levels. The transform procedure is designed for the particular class of images showing the head and shoulders of individuals facing the front, with a homogeneous background. A group of 12 passport photographs is used as a representative image set for testing purposes. Results obtained using the wavelet transform procedure are compared to the results of compressing the same images using the JPEG grayscale compression standard. The quality of the images is judged using the peak signal to noise ratio (PSNR). At compression ratios below 40:1, the JPEG compression system produced better quality images than the wavelet system. At compression rates above 40:1, the wavelet transform produced images of better quality. The wavelet transform produced images of above 30 dB PSNR at compression ratios of up to 150:1. These images are acceptable for subject recognition.