Adaptive Segmentation Models For Images With Intensity Inhomogeneity

Image segmentation is a fundamental problem in the field of computer vision and artificial intelligence,whose purpose is to extract objects to be processed in an image.Digital images are widely existed in many practical issues such as aerospace,remote sensing mapping,medical imaging,communications and transportation.Affected by the diversity of imaging equipment and other human factors,many images inevitably show the feature of inhomogeneous intensity.Up to now,many scholars have conducted a great deal of research aiming at the problem of how to accurately segment targets in uneven gray images.For many years,image segmentation methods based on partial differential equations have received extensive attention from scholars at home and abroad because of many advantages that traditional segmentation methods do not have.Geometric active contour models are known as one of successful methods.This dissertation focuses on geometric active contour models,for instance,the Region-Scalable Fitting(RSF)model and the Piecewise Constant Variational(PCV)model.With appropriate improvements on these two existing models,an adaptive segmentation model for images with intensity inhomogeneity is proposed.The main work of this article is as follows:(1)Given the problem that the contour initialization affect segmentation results of the RSF model,we propose an adaptive image segmentation model based on local neighborhood contrast.Firstly,the local and global fitting terms are constructed based on their image information respectively.Secondly,we quantify the change of local grayscale using the gray information of different neighborhoods of pixels,then an adaptive weight function adjusting the weight of two fitting terms is defined.Meanwhile,the L2 regularization term is selected to constrain the evolution of the contour.Finally,we raised an adaptive segmentation model with constant initialization.The adaptive weight function can assigns weights to the fitting terms,and the proposed model without initial contour can effectively segment various gray uneven images and noise images.(2)Due to the fact that the PCV model is not effective in processing images with intensity inhomogeneity and is sensitive to the initial contour,a grayscale uneven image segmentation model without re-initialization is proposed.This model combines the evolution equations of both the local image fitting(LIF)model and the PCV model.Finally,the model is obtained in the shape of partial differential equation.Experiments results show that the proposed model can not only enhances the robustness and accuracy of segmenting grayscale uneven images and noise images,but also can obtain stable segmentation results without many iterations.