Semi-interactive region segmentation based on sparse representation

Region segmentation is an important and challenging task. The applications range from tumour detection in medical imaging, computer aided surveillance, object location, pattern separation etc. Sparsity based data modelling in recent times have produced state of the art results in many image processing tasks. In this paper, we propose a semi-interactive region segmentation in sparse framework. Proper data modelling is the key to learning based segmentation. We propose a hybrid feature vector which is a combination of weighted RGB values and the proposed histogram estimated by first multiplying Gaussian weight to each count of the pixel intensity according to its respective position in the patch. We study the effect of various parameters such as patch size, number of atoms in dictionary, number of training feature vectors and sparsity constraint on the segmentation behaviour. We test our proposed segmentation algorithm on the subset of images from BSDS300 (Berkeley Segmentation Dataset).