OPTIMIZATION OF IMAGE SEGMENTATION THROUGH THE TRANSPORTATION PROBLEM FRAMEWORK

Abstract

Image segmentation is one of the key activities in image processing in which an image is divided into areas of significance to be analysed. Conventional approaches are based on clustering, thresholding or edge detecting which tend to be ineffective when noise is present, the illumination varies, or the object has complicated features. In this work, we suggest applying the model of optimal segmentation based on the use of the transportation problem which is a classical model of operations research. Segmentation is formulated as a cost minimization by considering pixel distributions as supply point and expected segment models as demand point. The model of transportation makes certain groups of pixels are optimally allocated to areas with insignificant spatial or intensity distortion. It uses the ideas of optimal transport theory and offers a strong and mathematically based theory of segmentation used in medical imaging, satellite image and pattern recognition.  Image segmentation is a critical part of image processing that focuses on the division of an image into meaningful components in order to analyse the same. The traditional techniques which aim at clustering, thresholding, or edge detection tend to fail in case of noise, or contrast in the lighting across the image, or complex topology. The new approach proposed by the current study is that of developing segmentation as a problem in transportation, which is a traditional model of operation research. In this model pixel intensities are distributed and supply nodes are modelled and the prototypes of the desired segment are desired demand nodes. The intention of this is to lessen the total transportation cost, distortion of space or intensity and therefore optimal distribution of pixels in regions is realized. The model is well entrenched in the optimal transportation theory; therefore, providing a mathematically sound and robust solution of the segmentation issues. Other restrictions that can be easily considered in the approach include spatial continuity or texture similarity to give it more flexibility. The experiments indicate that it is capable of doing well in many areas including medical imaging, satellite images analysis and pattern recognition with better segmentation accuracy and interpretability than the traditional methods.