Fast Two-stage Segmantation Based On Active Contour Model

Image segmentation plays an important role in both image analysis and computer vision.However,medical images,such as ultrasound,X-ray radiography/tomography and magnetic resonance imaging(MRI),are often distorted by complex noise and intensity inhomogeneity,which makes image segmentation research still be a challenging problem.In the past decades,a number of researchers have proposed many effective methods.Among them,the active contour model(ACM)has been proved to be a successful branch.In the paper,we developed the research into the method of image segmentation based on active contour model.Firstly,we summarize and classify the traditional active contour models.Following it,this paper focus on the mathematical background of region-based active contour model.Typical region-based models mainly aim to identify each region of interests by using a certain region descriptor to guide the motion of the active contour.Finally,we present the fast two-stage segmentation based on active contour models.The mainly research work and contribution of our model as follow:1.The local region-based methods have the capability of dealing images distorted by complex noise and intensity inhomogeneity.However,these models are sensitive to the initialization and takes a large number of iterations to converge.In this paper,inspired by the advantages of local region-based methods,we proposed a two-stage segmentation.The image is directly down-sampled in the first stage.And the local region-based method is employed to segment the down-sampled image.A coarse segmentation result is fast obtained with low computational complexity in this stage.The second stage employs an improved local region-base method.More importantly,the second stage uses the up-sampled coarse segmentation contour of the first stage as the initialization,which effectively avoids the difficulty about finding a suitable initialization for local region-based model.2.Due to information loss in the down-sampling operators,the coarse segmentation result is not accurate.The segmentation curve is shifted away from the real object boundary with an amount proportion.However,the up-sampling version of coarse segmentation result is very close to the real object boundary.Under the condition of the above fact,we improve the segmentation model with distance function.This method could not only correct the deviation between the second segmentation model and coarse contour,but also guarantee the stability of the curve evolution.3.Meanwhile,a fast technique based on Split Bregman method is introduced to effective solve the two segmentation problem,which is faster than gradient flow method solving PDEs to find the minimizers.Finally,a large number of simulation experiments are carried out using MATLAB software.In details,our segmentation process includes two steps.The first stage combines local region-based model with sampling methods,which reduce computation complexity effectively and provides a good initialization for next stage.In the second stage,up-sampling the final active contour of the first stage and directly used as the initial contour.Integrated with Split Bregman method,the local region-based model can be fast solved.Finally,the experimental results demonstrate our methods are robust to noise and rapidly given the accurate segmentation results.