Image segmentation algorithms based on the machine learning of features

For general purpose image segmentation, it is required to find and integrate the features that best characterize the regions to be segmented. This paper proposes a machine learning approach to finding the appropriate features and also a new segmentation method based on the information obtained while learning. Precisely, our method is based on the AdaBoost algorithm for learning the difference between regions, and the CRF-based (conditional random fields) energy formulation for the segmentation using the information from the learning. We have applied our method to interactive (semi-automatic) and unsupervised (fully-automatic) segmentation problems. While the interactive case is relatively straightforward due to the nature of our machine learning scheme, the unsupervised case is not. Hence, for the unsupervised segmentation, we devise a new initialization method and an EM-like (Expectation–Maximization) optimization method that iterates AdaBoost learning and graph-cuts. The analysis shows that the number of regions is automatically determined so that only distinguishable regions are survived. Experimental results also show that the proposed method gives promising results in diverse applications such as texture segmentation, color-texture segmentation, and page segmentation.