Local Dispersion Optimized LoG and Image Fitting for Intensity Inhomogeneous Image Segmentation

Region-based active contour models (ACM) have shown better performance compared to the edge-based models in numerous aspects like noise tolerance, inhomogeneity and background complexity correction, initialization robustness and speed of curve evolution. However, the fusion of edge and region based methods has circumvented the performance of the edge and region based methods working independently. Hence, this paper reveals a new technique to optimize Laplacian of Gaussian (LoG), the popular second order edge detector by local dispersion based edge estimation. The local dispersion computed image assists the optimization process to smooth the region without object boundary and simultaneously enhances the boundaries. The optimized LoG is added with the local image fitting (LIF) energy functional inside a variational framework with suitable scaling parameters. Thus an efficient fusion of region based energy and optimized edge energy performs with greater efficiency in terms of segmentation accuracy, noise tolerance, initial contour placement, iteration time and complex background suppression.

I. Introduction

Segmentation is the process of separating regions that are different in color, intensity and texture. In computerized image understanding and analysis, image segmentation acts as an intermediate step. Segmentation is a complex task for noisy and low-contrast images for lack of distinct image features. Furthermore, different sources of images make the task more challenging. In this context, various segmentation methods are developed which can be divided as classical methods and deep learning models. Both of them have their own advantages and short-comings. Deep learning models require huge set of training data to train a model to learn segmentation. This is time consuming and proper training data set may not be available for a specific type of image. In this regard, classical methods, being based on mathematical understanding and reasoning of images play an excellent role to process those images. Traditional image segmentation methods can be classified as, threshold-based, edge-based, geometric model-based, region growing based and active contour-based (ACM) [1]–[3].