Article Preview
TopIntroduction
As the intermediate steps, image segmentation is an important subject in machine vision (Rupak, Rama, & Garg, 2019; Robinson & Valindria, 2019; Chaki & Dey, 2019; Satapathy & Raja, 2018; Uhl & Leyk, 2018; Zhou & Wang, 2017). Its aim is to partition an image into several different parts according to its intensities, colors and shape, etc. However, due to the imperfection of imaging devices and complicated background, the phenomenon of intensity inhomogeneity, noise and weak edges often occur in medical images and natural images, which may affect the final segmentation result. So, it’s still a severe challenge for segmenting objects in clinical application. Over the past decades, extensive approaches have been wide application. Among the available schemes, level set algorithms attract considerable attention that is the most successful energy-based minimization framework. Generally, level set algorithms may be split into edge-based (Liu, Liu, & Xing, 2019; Zhi & Shen, 2018; Gupta & Anand, 2017; Ibrahim, Hasan, & Jalab, 2018; Cheng, Liu, & Xing, 2017) and region-based (Cheng & Tian, 2018; Cai & Liu, 2018; Meng & Wen, 2018; Liu & Fan, 2017; Li & Zeng, 2017).
Edge-based methods adopt the edge information of images as stopping criteria to guide contour curve approaching to the desired target edges. Thus, these methods can solve object with high contrast and have low computational cost. For example, the GAC algorithm, one of the most outstanding methods, is proposed by Caselles et al. in (Caselles, Kimmel, & Sapiro, 1997). However, without considering the gray distribution of local areas, it is difficult for the edge-based model to find the target. In (Li & Xu, 2005; Li & Xu, 2010), Li et al. constructed a classical DRLSE method with an energy penalizing which completely eliminated the re-initialization procedure. Nevertheless, these kinds of models cannot achieve satisfactory performance on images with low-contrast. Recently, various edge-based models that incorporate different image features have been proposed. In (Azizi, Elkourd, & Azizi, 2017), Azizi et al. presented a robust segmentation algorithm by utilizing the modified edge detector. Comparatively to ESF, this model can effectively solve the noise problem. Khadidos and Sanchez (2017) presented a method, which uses the local gradient features as weighting factor. As a consequence, this model lead to satisfactory segmentation results, particularly around weak edges. In (Peng & Ma, 2018), Peng et al. adopt the image edge information to construct the segmentation model. Based on a new diffusion with an edge indicator function, the proposed method can get relatively efficient with low contrast in images. In (Wang & Chen, 2018), Wang et al. construct an energy equation based on local edge entropy which can reduce unnecessary contour evolution and obtain the results efficiently. However, when the intensities of the images change gradually near the boundary, these methods usually fall into local minima. In (Pratondo, Chui, & Ong, 2015), Pratondo et al. utilize probability scores to build a new ESF equation, which can improve the effectiveness of traditional algorithm
Recently, non-local means filtering has been proved to be a popular algorithm to reduce noise while preserving edges. In this paper, the main innovation is to use fuzzy k-NN classification algorithm with non-local means filtering to construct level set function. The NLM filtering enhances the edges on the boundary and the fuzzy k-NN classification algorithm is more flexible since it utilizes the probability scores from any classifier. Hence, a unified framework using variational model is presented, which can address the limitations of low contrast and severe noisy boundaries. Main contributions of this method are:
1) The non-local means filtering is incorporated in a variational level set formula to acquire the edge graph. The noise interference is effectively reduced and more image information is obtained, which can be used for the subsequent external force field analysis. Therefore, this model has good robustness.