| Description |
Image segmentation is the problem of partitioning an image into disjoint segments that are perceptually or semantically homogeneous. As one of the most fundamental computer vision problems, image segmentation is used as a primary step for high-level vision tasks, such as object recognition and image understanding, and has even wider applications in interdisciplinary areas, such as longitudinal brain image analysis. Hierarchical models have gained popularity as a key component in image segmentation frameworks. By imposing structures, a hierarchical model can efficiently utilize features from larger image regions and make optimal inference for final segmentation feasible. We develop a hierarchical merge tree (HMT) model for image segmentation. Motivated by the application in large-scale segmentation of neuronal structures in electron microscopy (EM) images, our model provides a compact representation of region merging hypotheses and utilizes higher order information for efficient segmentation inference. Taking advantage of supervised learning, our model is free from parameter tuning and outperforms previous state-of-the-art methods on both two-dimensional (2D) and three-dimensional EM image data sets without any change. We also extend HMT to the hierarchical merge forest (HMF) model. By identifying region correspondences, HMF utilizes inter-section information to correct intra-section errors and improves 2D EM segmentation accuracy. HMT is a generic segmentation model. We demonstrate this by applying it to natural image segmentation problems. We propose a constrained conditional model formulation with a globally optimal inference algorithm for HMT and an iterative merge tree sampling algorithm that significantly improves its performance. Experimental results show our approach achieves state-of-the-art accuracy for object-independent image segmentation. Finally, we propose a semi-supervised HMT (SSHMT) model to reduce the high demand for labeled data by supervised learning. We introduce a differentiable unsupervised loss term that enforces consistent boundary predictions and develop a Bayesian learning model that combines supervised and unsupervised information. We show that with a very small amount of labeled data, SSHMT consistently performs close to the supervised HMT with full labeled data sets and significantly outperforms HMT trained with the same labeled subsets. |
