Instance segmentation and material classification in X-ray computed tomography

Abstract

Over the past thirty years, X-Ray Computed Tomography (CT) has been widely used in security checking due to its high resolution and fully 3-d construction. Designing object segmentation and classification algorithms based on reconstructed CT intensity data will help accurately locate and classify the potential hazardous articles in luggage. Proposal-based deep networks have been successful recently in segmentation and recognition tasks. However, they require large amount of labeled training images, which are hard to obtain in CT research. This thesis develops a non-proposal 3-d instance segmentation and classification structure based on smoothed fully convolutional networks (FCNs), graph-based spatial clustering and ensembling kernel SVMs using volumetric texture features, which can be trained on limited and highly unbalanced CT intensity data. Our structure will not only significantly accelerate the training convergence in FCN, but also efficiently detect and remove the outlier voxels in training data and guarantee the high and stable material classification performance. We demonstrate the performance of our approach on experimental volumetric images of containers obtained using a medical CT scanner.