An Efficient Interactive Multilayered Approach for Segmentation of Complex Images

This paper presents an interactive algorithm that is fast and accurate for segmentation of multiclass images called GMMLC (Gaussian Mixture Model in Laplacian Coordinates). The proposed framework has an imperative feature of classification of complex images with similar background and foreground features. This paper refers to such images as Indistinct and Indistinguishable Boundary images. GMMLC combines the prevalent features of the multivariate Gaussian Mixture Model and the weighted valency graph based approach of Laplacian Coordinates. Commendable results are obtained by combining an efficient but suboptimal solution defined in probability space and the average neighborhood preservation properties of LC. This work is inspired by Casca’s seed based image segmentation approach. However, this technique presents significantly improved results in terms of boundary fitting and efficiency with only a small amount of seeds due to the homogenous regions obtained from GMM preprocessing. The proposed framework outperforms all existing advanced techniques by means of greater efficiency and better boundary fitting with minimum user intervention. The algorithm is applied on 500 images from the Berkeley datasets (BSD500 and BSD300) and 50 images from the Grab cut dataset (PASCAL VOC 2007). The paper culminates with a qualitative boundary and efficiency evaluation in comparison with five prominent algorithms BPT, SRG, GC, SIOX and LC.