Object detection in biological images using a search-based framework

Graduation date: 2014Access restricted to the OSU Community, at author's request, from June 20, 2014 - June 20, 2015This thesis addresses a basic problem in computer vision, that of semantic labeling of images. Our work is aimed at object detection in biological images for evolutionary biology research. In particular, our goal is to detect nematocysts in Scanning Electron Microscope (SEM) images. This biological domain presents challenges for existing approaches developed to address other domains (e.g. natural scenes). An image may show more than one nematocyst under partial occlusion and amidst background clutter (e.g., cellular debris). We formulate the detection of nematocysts as labeling of a regular grid of patches, where patches occupied by a nematocyst are labeled as foreground and the other patches are labeled as background. This structured prediction problem is addressed using the $\mathcal{HC}$-Search framework. $\mathcal{HC}$-Search seeks a solution in the space of candidate labelings of image patches. It employs a heuristic function ($\mathcal{H}$) to uncover high-quality candidate labelings, and then applies a cost function ($\mathcal{C}$) to select the best-scoring prediction among the candidates. The heuristic and cost functions are learned on training data using imitation learning. Our key novelty is a formulation of a stochastic search space and pruning function that improve the accuracy and efficiency of $\mathcal{HC}$-search. The stochastic search space is generated by a successor function that proposes random candidate labelings based on clusterings of patches in the image instead of employing the traditional method of flipping the labels of individual patches. Additionally, we introduce a pruning function ($\mathcal{P}$) to remove bad candidate labelings generated by the successor function. This induces a sparser search space, which improves the efficiency of search. We compare our approach with the Conditional Random Field (CRF) model, a well-understood framework for scene labeling. While CRF inference typically yields good performance on natural scenes, our results demonstrate that CRFs perform poorly on the nematocyst images and that $\mathcal{HC}$-Search outperforms CRFs