Treating images as functions and using variational calculus,
mathematical imaging offers to design novel and continuous methods, outperforming traditional methods based on matrices, for modelling real life tasks in image processing.
Image segmentation is one of such fundamental tasks as many application areas demand a reliable segmentation method. Developing reliable selective segmentation algorithms is
particularly important in relation to training data preparation in modern machine learning as accurately isolating a specific object in an image with minimal user input is a valuable tool. When an image's intensity is consisted of mainly piecewise constants, convex models are available.
Different from previous works, this paper
proposes two convex models that are capable of segmenting local features defined by geometric constraints for images having intensity inhomogeneity.
Our new, local, selective and convex variants are extended from the non-convex Mumford-Shah model intended for global segmentation.
They have fundamentally improved on previous selective models that assume intensity of piecewise constants. Comparisons with related models are conducted to illustrate the advantages of our new models.