Active Contour Model Driven By Pre-fitting Bias Field For Image Segmentation

Object identification and feature analysis in image are directly affected by image segmentation,which is a key pre-treatment in digital image processing.Due to the importance of image segmentation,many researchers have proposed an array of algorithms to segment object regions in the image.Among these methods,the complete variational theory and good segmentation performance has made active contour model become hotspot for the researchers.However,there exist intensity inhomogeneity(also called bias field)in digital images due to imperfect imaging device,uncertain light condition and complex background information.Traditional image segmentation methods are prone to bias field,which leads to incorrect image segmentation.To solve this problem,a novel active contour model driven by pre-fitting bias field has been proposed.Different from traditional active contour model,the proposed model utilizes bias field and gradient information to formulate energy functional,which is good at detecting the contour of target object in the image.The main contributions in his paper are as follows:(1)Optimize the method of bias field computation.According to assumptions of bias field,local grayscale mean algorithm is adopted to approximate bias field.Local grayscale mean algorithm has low computational complexity and needs little CPU tine to approximate bias field.Experiments validate that local grayscale mean algorithm not only can achieve bias field correction,but also can provide a foundation for following image segmentation.(2)Formulate energy functional based on pre-fitting bias field.Pre-fitting bias field is utilized to construct energy function in active contour model.Since the bias field computation is implemented before iteration,and thus it is unnecessary to update bias field during iteration.At the same time,the curve evolution efficiency also gets improved..(3)Insert new adaptive edge indicator function based on the energy function driven by pre-fitting bias field.The improved energy function makes curve evolution fast when the evolving contour is placed far away from object boundary,and makes curve evolution slow when the evolving contour approaches the object boundary.Accordingly,the high segmentation accuracy and efficiency can be guaranteed.(4)Optimize gradient descent equation and the method of regularizing level set.The improved gradient descent equation not only limits the magnitude and slope of data driven term,but also improves energy function curve in iterative process.Therefore,the stability of curve evolution is ensured.Besides,the proposed regularization helps level set maintain the signed distance property and avoid wrong curve evolution,which ensures the advantage over segmentation speed and precision.Experiments on different kinds of test images show that the proposed model is competent to overcome intensity inhomogeneity,and has good robustness to initial contour and noise.