Affiliations: Computational Intelligence Group, University of the Basque Country, UPV/EHU, Bilbao, Spain
Correspondence:
[*]
Corresponding author: Manuel Graña, Computational Intelligence Group, University of the Basque Country, UPV/EHU, Bilbao, Spain. E-mail: [email protected]
Abstract: This paper introduces the Hybrid Extreme Rotation Forest (HERF) classifier describing two succesful applications in the image segmentation domain. The HERF is an ensemble of classifiers composed of Extreme Learning Machines (ELM) and Decision Trees. Training of the HERF includes optimal rotation of random partitions of the feature set aimed to increase diversity. The first application is the segmentation of 3D Computed Tomography Angiography (CTA) following an Active Learning (AL) strategy for the optimal sample selection to minimize the number of data samples needed to obtain a required accuracy degree. AL is pertinent for interactive learning processes where a human operator is required to select training samples to enhance the classifier in an iterative process, therefore labeling samples for training may be a time consuming and expensive process. CTA image segmentation is one of such processes, due to the variability in CTA images which hinders the generalization of classifiers trained on one dataset to new datasets. Following an AL strategy, the human operator is presented with a visual selection of pixels whose labeling would be most informative for the classifier. After adding those labeled pixels to the training data, the classifier is retrained. This iteration is repeated until image segmentation quality meets the required level. The approach is applied to the segmentation of the thrombus in CTA imaging of Abdominal Aortic Aneurysm (AAA) patients, showing that the structures of interest can be accurately segmented after a few iterations using a small data sample. The second application is a new semisupervised segmentation algorithm for hyperspectral images. The algorithm steps are: 1) supervised training an initial classifier from a small balanced training set, 2) clustering of the image pixels, by a k-means algorithm 3) adding unlabeled pixels to the original trainning data set according to the spatial neighborhood and the cluster membership, 4) supervised training of the classifier with the enriched training data set, 6) classification of the hyperspectral image 4) spatial regularization of classification results consisting in selecting the most frequent class in each pixel neighborhood. Results on two well known benchmarking hyperspectral images improve over state of the art algorithms.
