Robust Learning and Segmentation for Scene Understanding

摘要

This thesis demonstrates methods useful in learning to understand images from only a few examples, but they are by no means limited to this application. Boosting techniques are popular because they learn effective classification functions and identify the most relevant features at the same time. However, in general, they overfit and perform poorly on data sets that contain many features, but few examples. A novel stochastic regularization technique is presented, based on enhancing data sets with corrupted copies of the examples to produce a more robust classifier. This regularization technique enables the gentle boosting algorithm to work well with only a few examples. It is tested on a variety of data sets from various domains, including object recognition and bioinformatics, with convincing results. In the second part of this work, a novel technique for extracting texture edges is introduced, based on the combination of a patch-based approach, and non-parametric tests of distributions. This technique can reliably detect texture edges using only local information, making it a useful preprocessing step prior to segmentation. Combined with a parametric deformable model, this technique provides smooth boundaries and globally salient structures.