Learning robust object segmentation from user-prepared samples

Object segmentation is a crucial task in image processing. A robust semi-automated object segmentation approach is introduced in this paper. The proposed approach learns segmentation from a small number of user-prepared (ground-truth) samples. The segmentation is performed in two main sequential stages, namely, target object localization by applying optimal mathematical morphology procedure, and segmentation, by conducting some basic image processing operations. The outstanding feature of this approach is, unlike other existent approaches, that it does not need apriori knowledge or a large number of samples to learn from. The training is performed for a group of images to segment a specific object. The performance of the approach has been examined by a comprehensive well-designed validation set. For all test images, the target object was segmented accurately and the conducted experiments clearly show that the proposed segmentation approach is highly invariant to noise, rotation, translation, overlapping, scaling, and combination of them. The architecture of the approach and employed methodologies are explained in detail. Results are provided.