Schedule

Next Talk

Date: Monday, February 25th, 2013, 1-2pm

Location: Banting Institute, Lecture Hall Rm 131, 100 College Street, Toronto, Ontario (map)

Presenter: Dominique Brunet, University Health Network

Title: A Study of the Structural Similarity Image Quality Measure with Applications to Image Processing

Description: Image quality assessment is an essential step for the validation of many image processing algorithms (e.g. compression, restoration, interpolation, fusion, registration and, to some extent, segmentation). Ideally, a human observer would need to subjectively validate the performance of an algorithm by looking at samples of processed images. In practice, this proves too costly, particularly when the observations are made by one or several experts. Also, this kind of procedure is not amenable in an optimization framework where one wants either to tune some parameters or, better, to design a new image processing algorithm.

Traditionally, the Mean Square Error (MSE) was used as an objective measure of quality despite the well-known fact that it correlates poorly with subjective data. More recently, the Structural Similarity (SSIM) index has been introduced and gained widespread popularity in the image processing community.

In this talk, a study of the SSIM index undertaken during my doctoral years will be presented. The main questions to be answered are if, when, how and why the SSIM index is a valid quality measure for image processing tasks. Two points need to be addressed in order to answer of these questions: 1) the amenability of the SSIM index for various image processing applications, and 2) the validity of the SSIM index itself, particularly for medical images.

Specifically, I will first present my work on the properties of the SSIM index and of its generalization into normalized metrics. Then, I will briefly show some SSIM-optimal solutions for various settings relevant to image processing: affine transforms, best approximations for orthogonal or redundant bases, projections onto a convex set, geodesics and point estimators. Finally, I will outline recent experiments that were done in order to validate the SSIM index as a predictor of quality for lossy medical image compression.