Global optimization in discretized parameter space for predefined object segmentation
Huy Hoang Nguyen, Hyuk-Ro Park, Joon Seub Cha, Le Thi Khue Van, Guee-Sang Lee · 2013
In object segmentation field, while the non-predefined object segmentation distinguishes arbitrary self-assumed object from background, predefined object segmentation pre-specifies object evidently. This paper presents a new method to segment predefined objects by globally optimizing an orientation-based objective function that measures the fitness of object boundary in a discretized parameter space. A specific object is explicitly described by normalized discrete sets of boundary points and corresponding normal vectors with respect to its plane shapes in a certain aspect. The orientation factor provides robust distinctness for target objects. By considering the order relation of transformation elements, and their dependency on derived over-segmentation outcome, the domain of translations and scales is discretized efficiently. The appropriate transformation parameters of a shape model corresponding to a target object in an image are determined using the global optimization algorithm branch-bound. Discrete boundary points of the consequent transformed model are chained together to produce the final contour of the target object. The results tested on PASCAL dataset show a considerable achievement in solving complex background and unclear boundary images.