Graph Based Semi-supervised Learning in Computer Vision

OF THE DISSERTATION Graph Based Semi-Supervised Learning in Computer Vision by Ning Huang Dissertation Director: Joseph Wilder Machine learning from previous examples or knowledge is a key element in many image processing and pattern recognition tasks, e.g. clustering, segmentation, stereo matching, optical flow, tracking and object recognition. Acquiring that knowledge frequently requires human labeling of large data sets, which can be difficult and time-consuming to obtain. One way to ameliorate this task is to use Semi-supervised Learning (SSL), which combines both labeled and raw data and incorporates both global consistency (points in the same cluster are likely to have the same label) and local smoothness (nearby points are likely to have the same label). There are a number of vision tasks that can be solved efficiently and accurately using SSL. SSL has been applied extensively in clustering and image segmentation. In this dissertation, we will show that it is also suitable for stereo matching, optical flow and tracking problems. Our novel algorithm has converted the stereo matching problem into a multi-label semi-supervised learning one. It is similar to a diffusion process, and we will show our approach has a closed-form solution for the multi-label problem. It sparks a new direction from the traditional energy minimization approach, such as Graph Cut or Belief Propagation. The occlusion area is detected using the matching confidence level, and solved with local fitting. Our results have been applied in the Middlebury Stereo