JPEG Compression in the Intensity Domain

The JPEG compression scheme converts an image from RGB color coordinates (appropriate for display) to YCrCb coordinates (dividing the image into perceptually significant color and black-and-white image planes). The JPEG compressor divides each image plane into 8x8 blocks and performs the discrete cosine transform (DCT) to each block. It then quantizes the DCT coefficients based on how easily perceived each spatial frequency is: higher spatial frequencies get quantized more roughly than DC and near-DC components. These quantized coefficients are stored more compactly than the original coefficients or image, and remains more perceptually similar to the original than an equal compression scheme in the spatial domain.

Each of the steps in the JPEG compression scheme in linear and invertible except for the quantization step. This step results in a fair preservation of the mean levels of the framebuffer values of the image. but the framebuffer values are not linearly related to intensity. So after quantization, high-frequency variations have been averaged to their mean framebuffer value, which may correspond to a lower mean intensity value than the original. Conceptually we see that this difference will be more noticeable when the intensity values are higher to begin with, because this is where equally spaced framebuffer values correspond to instencity values of increasingly greater increments.

The purpose of this project is to study the mean-level distortion effects of JPEG compression, and propose and evaluate a scheme that compresses images in a linear intensity domain.

Graylevel

Intensity

Click here to continue.