Initial Observations and Proposed Compression Scheme

To confirm out hypothesis that distortions in level intensity is more noticeable when the mean level of the original image is high to begin with, we create test images of varying spatial frequencies. In this cae, I used a sinusoid whose frequency increased from left to right. Below are two images, one with full range of contrast, the other with grayscale framebuffer values ranging from .8 to 1.

Full contrast

Full contrast, JPEG compressed

Low contrast, high mean level.

Low contrast, high mean level, JPEG compressed

Proposed Compression Scheme

The JPEG compression, as mentioned before, occurs in the YCrCb coordinates. The Y values are similar to the L* values in the CIELAB and CIELUV color standards, in that each step in the Y scale corresponds to an equal increment in perceived brightness. This is approximated by Steven's Power Law and is an non-linear space. The immediately obvious thing to do then is to try converting the Y values (from the YCrCb coordinates) to true Y intensity values. In this new scale each step corresponds to an equal increment of light intensity. In this linear space, mean levels would be preserved. The intensity values would have to be converted back to framebuffer values after decompression.

There are some practical problems with this. First, we lose all the benefits of of the perceptually linear space, such as automatic gamma correction, noise minimization, and efficient use of bits. Furthermore, coding in a peceptually non-linear space suggests that we will create quantization artifacts. But we will implement it anyway and study the potential benefits that such a scheme offers. Below are the two test images introduced above, and the their compressed form under both the regular JPEG schemes, and the JPEG compression in intesnity space, which we shall call Y-JPEG for brevity. We see that indeed we get coarse quantization levels and concomitant artifacts at low intensities, but that at high intensities we get better compression. Now we will implement the the scheme for color images.

Full contrast

Full contrast, JPEG compressed

Full contrast, Y-JPEG compressed

Low contrast, high mean level.

Low contrast, high mean level, JPEG compressed

Low contrast, high mean level, Y-JPEG compressed

Click here to continue.

If you have questions or comments, please e-mail Alan Tseng / alant@stanford.edu.