PROJECT TITLE :

A Spatially Constrained Probabilistic Model for Robust Image Segmentation

ABSTRACT:

In probabilistic model based segmentation, the hidden Markov random field (HMRF) is used to describe the class label distribution of an image. In the current HMRF models, either the number of surrounding pixels with similar class labels or the spatial distance of neighbouring pixels with differing class labels are used to determine the distribution of class label distributions. Furthermore, this spatial information is solely used to estimate the picture pixel's class labels, while its contribution to parameter estimation is completely ignored. As a result, parameter estimation suffers, leading to subpar segmentation results. As a result of this, pixels in image boundary regions are frequently mislabeled in existing models since the models use the image's spatial information equally for all pixel label estimations. Accordingly, a new clique potential function and a new distribution of class labels are developed in this research and include data from picture class parameters. This framework includes a new scaling parameter that evaluates the contribution of spatial information for class label estimation of image pixels, unlike conventional HMRF model-based segmentation techniques. Changes in HMRF-based segmentation approaches demonstrate the importance of this framework. The advantages of the suggested class label distribution can be seen regardless of the intensity distributions. A comparison of the proposed and existing HMRF model class label distributions for brain MR image segmentation, HEp-2 cell delineation, and natural image and object segmentation shows that the suggested distributions outperform the existing distributions both qualitatively and numerically.