Active Contours Driven By Local Entropy Fitting Energy For Image Segmentation With Intensity Inhomogeneity

Image segmentation is a very important research topic in the field of computer vision.Image segmentation is used to locate objects and boundaries,its purpose is to simplify or change the representation of the image,make the image easier to understand and analyze.It has been widely used in the target recognition,moving object tracking,biomedical image analysis and other fields.One of the core challenges of image segmentation is the inhomogeneity of image grayscale caused by imaging equipment,lighting conditions and complex backgrounds.Although the existing algorithms can segment the target object well in the homogenous image,they have made errors in the inhomogeneous gray scene.Since the inhomogeneous gray images are more common in nature,the accurate segmentation of inhomogeneous objects is of great significance.To solve these problems,this thesis proposes a new image segmentation algorithm based on local entropy active contour model.By combining the global information and local information of the image,the image information can be used more comprehensively to make the segmentation result better.The main innovations of the model in this thesis are as follows:First,in many current models,Gaussian kernel function is used to constrain the range of local regions,but local information of different regions in the whole image is different.A single Gaussian kernel coefficient constrains the same size of the local area range,resulting in defects and errors in the evolution process.In this thesis,the model uses the local image entropy value to define the multi-Gaussian kernel coefficient.According to the local area of the image in different positions,the corresponding Gaussian kernel coefficient makes the division of the local area more reasonable,and effectively enhances the segmentation accuracy of the model for the inhomogeneous gray image.Secondly,construct an adaptive combination coefficient function according to the evolution process of the model,which divides the image segmentation process of the model in this thesis into two stages: the first stage,the global fitting item is the main part of the model,driving the level set function to evolve the approximate contour range of the target object;the second stage,the weight of the local fitting item increase,and then finely segment the edge details of the target object.The staged evolution through adaptive coefficients can reduce the evolution error of the model and achieve the best evolution result.Thirdly,based on the length energy term of the traditional active contour model,a new edge detection factor is constructed and added.The improved length term can slow down and prevent the evolution curve contour from approaching the actual target object,accelerate the evolution rate of the active contour model,and enhance the ability of the model to deal with inhomogeneous gray images.Based on the above innovation points,a new active contour model was established,and the proposed model can effectively segment the target object in the inhomogeneous image.Experimental results show that the proposed model has good robustness to external noise and initial area.Compared with the existing mainstream active contour model,the proposed model can detect the details of the edge of the object more accurately in the comparative experiment of segmenting inhomogeneous images such as medical images and outdoor images.It not only ensures the segmentation accuracy of the target object image,but also prevents over-sensitivity to the boundary region,and the results of our model detection are more reasonable and accurate.