**Contact Info**

Department of Applied Mathematics

University of Waterloo

Waterloo, Ontario

Canada N2L 3G1

Phone: 519-888-4567, ext. 32700

Fax: 519-746-4319

PDF files require Adobe Acrobat Reader

Visit our COVID-19 information website to learn how Warriors protect Warriors.

Please note: The University of Waterloo is closed for all events until further notice.

Wednesday, August 12, 2015 — 9:30 AM EDT

MC 5479

Daniel Otero | Applied Math, University of Waterloo

Function-valued Mappings and SSIM-based Optimization in Imaging

In image processing there are two important paradigms upon which imaging tools are developed: representation of images as either real or valued-vector functions, and using the Euclidian distance as either a metric or a norm for images. The former offers a good framework for applications since some type of images, e.g., colour and hyperspectral images, can be well represented by a vector-valued function. As for the latter, this is a convex and mathematically tractable metric that is frequently used in a variety of imaging tasks in which an optimization problem is to be solved (e.g., deblurring, denoising, sparse reconstruction, etc.).

Although many efficient imaging tools have been developed which are based on these paradigms, these approaches have some limitations. For instance, consider the case of a video sequence, which can be represented as a vector-valued function *u*(*x*) : *Ω* *→* *R ^{n}*. Here, Ω is the spatial domain, and each component

As for the Euclidian distance, measures of visual quality such as the Mean Square Error (MSE) and Peak to Signal Noise Ratio (PSNR) are Euclidian-based metrics, however, it has been shown that these metrics are not the best choice when it comes to measure the visual quality of images. This drawback of Euclidian-based metrics has been overcome up to some point by the Structural Similarity Index Measure (SSIM), which has shown to be a good model of the Human Visual System (HSV) [3]. For this reason, the SSIM has been employed as a fidelity term in several types of optimization problems for carrying out a variety of imaging tasks—e.g., filtering, best-SSIM approximation, video coding, etc. [1, 2, 4]. Nevertheless, these methods tend to be developed for particular applications and are not general.

In this thesis we propose two alternative approaches to overcome the difficulties mentioned above, namely, *Function-valued Mappings *(FVMs) and *SSIM-based Optimization*. With FVMs we address the problem of modelling complex data sets properly. In this case, we define a FVM as follows:

*u(x) : X **→* *F(Y),*

where both *X *and *F(Y)* are Banach spaces, and *F(Y)* is a space of real-valued functions whose support is the set *Y *. Observe that the range of *u *is infinite dimensional, therefore, depending on the application, an appropriate space *F(Y)* can be chosen to describe the functions that are assigned to each *x *∈ *X*. For instance, a hyperspectral image can be represented as a FVM of the form *u*(*x*) : *X* *⊂**R ^{2}*

Regarding SSIM-based optimization, we present a general framework for optimization problems that involve the SSIM as a fidelity term. Such a framework is divided into two approaches, namely, *unconstrained and constrained SSIM-based optimization*. In the unconstrained approach, an optimization problem is defined as follows:

*min {T (**Φ**(x), y) + **λ* *h(x)}, *

where *T *(·,·) := 1−SSIM(·,·) is a SSIM-based dissimilarity measure, Φ is a linear transformation, *y *a given observation, and *h*(*x*) a convex regularizing term along with its regularization parameter λ . As for the constrained counterpart, the minimization problem is given by

*min T (**Φ**(x), y)*

*subject* *to h _{i}(x)*

*Ax=b, *

where the *h _{i}*(

**References**

[1] D. Brunet, E.R. Vrscay and Z. Wang, “Structural similarity-based approximation of signals and images using orthogonal bases,” in ICIAR 2010, Vol. 6111 of Lecture Notes in Computer Science, Springer, 11–22 (2010).

[2] S.S. Channappayya, A.C. Bovik, C. Caramanis, R.W. Heath, Jr., “Design of linear equalizers optimized for the structural similarity index.” IEEE Transactions on Image Processing, 17(6), 857–872 (2008).

[3] Z. Wang, A.C. Bovik, H.-R. Sheikh and E.S. Simoncelli, “Image quality assessment: From error visibility to structural similarity.” IEEE Trans. Image Processing, 13(4), 600–612 (2004).

[4] S.Wang,A.Rehman,Z.Wang,S.MaandW.Gao,“SSIM-motivated rate-distortion optimization for videocoding.” IEEE Trans. Circuits Syst. Video Techn., 22(4), 516–529 (2012).

**Contact Info**

Department of Applied Mathematics

University of Waterloo

Waterloo, Ontario

Canada N2L 3G1

Phone: 519-888-4567, ext. 32700

Fax: 519-746-4319

PDF files require Adobe Acrobat Reader

University of Waterloo

University of Waterloo

43.471468

-80.544205

200 University Avenue West

Waterloo,
ON,
Canada
N2L 3G1

The University of Waterloo acknowledges that much of our work takes place on the traditional territory of the Neutral, Anishinaabeg and Haudenosaunee peoples. Our main campus is situated on the Haldimand Tract, the land promised to the Six Nations that includes six miles on each side of the Grand River. Our active work toward reconciliation takes place across our campuses through research, learning, teaching, and community building, and is centralized within our Indigenous Initiatives Office.