ABSTRACT
This chapter explores techniques and methodologies involved in formulating the problem through frameworks involving classification techniques, regression techniques, and combinations thereof. It describes nonlinear regression algorithms whose parameters have been primarily determined through relationships learned from a training set. Pattern recognition and machine learning define the act of taking data collected a priori, observing relationships inside the data, and generalizing the learned relationships. The attribute that distinguishes pattern recognition techniques from other types of algorithms is a training set. Support vector regression (SVR) is especially useful to image super-resolution because it can directly model and represent the highly nonlinear and complex relationship between lowand high-resolution image data. Consequently, the sheer number of parameters for an SVR with an imperfect kernel space may render the optimization problem unmanageable. Diversity and complexity in image patches suggest that super-resolution is inherently a nonlinear operation.
