Statistical Based Image Interpolation

Superresolution, or image upscaling, is the process by which an image is spatially enlarged or brought to higher resolution. Simpler techniques for superresolution such as bilinear and bicubic interpolation do not consider any information but that which is provided by the low-resolution image that is to be enlarged. The resulting image from these techniques is often blurry and do not improve in resolution much. Those techniques that do appear to perform well usually approach the superresolution problem through an optimization of visual acceptableness without consideration of the actual accuracy of the resulting image. Because the measure of the image correctness has generally been accepted as some form of mean squared error evaluation, it is only prudent to estimate these images to optimize this criterion. Algorithms that perform considerably better in this respect are those that use statistical learning because they bring to the table additional information that has been determined a priori. Within this category of algorithms, images can be evaluated and altered in a few domains, mainly spatial and frequency domains. The majority statistical learning algorithms when applied to superresolution work in the spatial domain. There are some shortcomings of working in the spatial domain, so a novel approach to the superresolution problem is introduced in the frequency domain, which uses an extension to the Decimation in Time property of the Two-Dimensional Discrete Cosine Transform. With this property, the relationship between a lowresolution image block and its upscaled equivalent can assume a structure. This structure should aid in the application of statistical learning to superresolution. There are several choices when dealing with statistical learning. We use Support Vector Machines in the form of regression and classification because of its optimality and speed. Using Support Vector Machines, information from a training set will allow a feature set consisting the Discrete Cosine Transform of a lowresolution image to approximate the high-resolution image.