Tuesday, February 22nd, 2005
I’m currently taking a course on digital video processing given by Prof. Thrasyvoulos Pappas, my advisor in the Image and Video Processing Laboratory (IVPL) at Northwestern.
For the course project, I’m studying objective image quality metrics, or the computation of a number that corresponds to the perceived quality of an image.
One image quality metric that is often used when comparing a reference and degraded image is the mean squared error (MSE), computed by simply averaging the squared differences between the reference and degraded image. For example, the degraded image could be a highly compressed version of the reference. While MSE is simple to understand and easy to compute, it does not achieve a good correspondance with perceived image quality.
Some interesting image quality methods have been proposed and tested recently. Junquig Chen from the IVPL evaluates metrics used when optimizing image compression, comparing MSE with subband, wavelet, and DCT-based metrics (see the SPIE paper).
Also, some very intersesting work has come from Eero Simoncelli’s Laboratory for Computational Vision (LCV) at New York University. Zhou Wang’s work on his Structural SIMilarity (SSIM) index is the best approach I’ve found so far for quantitatve evaluation of image quality for many different applications.
In upcoming blog entries, I hope to summarize and review some of the most interesting and influential papers that deal with image quality. I’ll start with Zhou Wang’s “Image Quality Assessment: From Error Visibility to Structural Similarity”. Stay tuned ….
