PDE Image Segmentation Based On Semi-Local Region Information

Image segmentation not only plays an important role in the image processing, but also acts as the key step in computer vision. The stand or fall of partition results directly influence the subsequent processing of the high-level image analysis and understanding. Up to now, a variety of algorithms have been proposed to solve the image segmentation problems. However, most of them are sensitive to noise and appear discontinuous edge easily. The partial differential equations (PDE) based image processing method stands out with its high accuracy and keeping edge continuityIn this paper, the Snake model, M-S model, C-V model, LBF model, LIF model and Sub-Local model are introduced in details from the two aspects of active contour based on boundary and active contour based on region. Moreover, the advantages and disadvantages of each model are elaborated.Then aiming at the fact, three kinds of image segmentation methods based on the semi-local region information are proposed in this paper for the images that the background areas distribute uniformly and object areas distribute disorderly. The three models inherit the advantages C-V model and LBF model. At the same time, they abandon the drawbacks of C-V model which only applies in the images with intensity homogeneity and one-phase LBF model which doesn’t work for the images with multiple junctions. In addition, it also overcomes the drawbacks of the Sub-Local model. Firstly, a novel semi-local region based image segmentation model named LBF&CV_B is proposed. The LBF& CV_B model employs the global energy and local energy to approximate the intensities of the objects. And it utilizes a piece-wise smooth function to depict the intensities of background. Furthermore, a balloon term is added to increase computation efficiency. Experiments show that the LBF&CV_B model gets better performance than C-V and LBF model no matter for the homogeneous and inhomogeneous images or some part of noise images. Because the evolution curve of LBF&CV_B model can’t converge to the real boundary finally, the LBF&CV_GAC model is proposed. To compare with LBF&CV_B model, it utilizes weighted arc length of generalized GAC to substitute the balloon term and regularized term that makes the evolution curve more stable. The experimental results indicate that LBF & CV_GAC model plays very well than others. And the convergence of LBF&CV GAC model is better than LBF&CV B model and Sub-Local model. In the mean time, it will not appear the oscillation phenomenon which the LIF model and Sub-Local model may occur. While further researches expose that the LBF & CV_GAC model evolves slower. Meanwhile, it always fails in segmenting the images with smooth background and deep recessed area with small interval. Thus the ILBF & CV_GAC model is proposed. By adjusting the scale parameters of inside and outside the curve, it speeds up the evolution of ILBF & CV_GAC model and can segment some other images that LBF & CV_GAC model can’t.