Stochastic Optimization for Learning-based Super-resolution: Algorithms and Applications
Human beings get much of their information visually and depend on perception of images for many critical tasks, such as object identification, medical image analysis, photography, etc. In many visual-based applications, higher resolution images are required for perceiving and receiving critical information. A high resolution image can contribute to a better identification of a suspect’s face, or a more accurate localization of a tumor in a mammogram, or a more pleasing view in high definition television, and so on. However, it is hard to obtain the high resolution images needed for some applications, for example, the cost of sensors increases as a exponential function of their resolution so that a high resolution sensor may be too expensive. Another example is in X-ray imaging for medicine. Higher incident X-ray beam intensity produces higher resolution images but harms patients. In some cases image resolution can be improved through image super-resolution, an image processing procedure that takes a degraded image or a sequence of images as input and produces an image or a sequence of images of higher quality as output. This dissertation addresses the problem of optimization of super-resolution algorithm according to the specific requirements of the applications for which the images are used. One application addressed in this thesis is super-resolution for surveillance videos. In surveillance applications, cameras are usually set up with wide fields of view to capture as much of the scene as possible. This normally results in low-resolution images of suspects’ faces, therefore face image super-resolution is critical for face recognition tasks. This thesis improves the efficiency of face image super-resolution using stochastic search for local modeling by exploiting the fact that face patches maintain relatively tight distributions for shape at successive iterations. Furthermore, though face super-resolution could improve the appearance of face images dramatically, the detailed facial features such as eyes, eyebrows, nose, mouth and teeth of the super-resolution face are different from the ground truth. This thesis studies whether face superresolution can help face identification by either human beings or computers, and proposes a novel patch-based simultaneous face identification and super-resolution approach that integrates face identification and super-resolution together. Another application addressed in this thesis is super-resolution for medical imaging. The high-quality mammogram is the most effective technology presently available for breast cancer screening. This thesis studies mammogram super-resolution, or synthesizing a high-resolution mammogram from an input low-resolution mammogram. Additionally, as medical imaging moves towards complete digital imaging and produces very large amounts of data, compression is necessary for storage and communication purposes. Super-resolution has the potential to eliminate the need to store images at full resolution, since they could be re-generated from low-resolution ones. This thesis explores mammogram compression using super-resolution, and then proposes a novel hybrid compression method for mammograms using super-resolution algorithms as a post-processing of the compression process. To further determine whether compression affects clinical diagnostic performance, this thesis studies whether these differences would affect micro-calcification detection by applying an computer-aided micro-calcification detection system to the original images and compressed images respectively, and then comparing the detection rates. Lastly, this thesis considers super-resolution for 3D face scanning. Detailed facial geometry contributes significantly to the visual realism of face models in computer games, movies, virtual reality applications, and so on. This thesis proposes a new technique for real-time high-resolution 3D facial scanning using stochastic super-resolution to generate a specular normal map based on the diffuse normal map, instead of capturing both of them during scanning process. This thesis also demonstrates that the technique can be also used to transfer specular normal information to unpolarized data, or transfer specular normal information from one person to other people, or the synthesized normal map can be used for image-based lighting.
Zheng, Jun, "Stochastic Optimization for Learning-based Super-resolution: Algorithms and Applications" (2010). ETD Collection for University of Texas, El Paso. AAI3443255.