广义核或混合核FLICM畜肉图像分割方法

Meat Image Segmentation Using Fuzzy Local Information C-Means Clustering for Generalized or Mixed Kernel Function

针对传统核模糊C均值聚类(Kernel Fuzzy C-Means,KFCM)畜肉图像分割方法对噪声适应能力不强的问题,提出基于广义核函数或混合核函数的模糊局部信息C均值聚类(Fuzzy Local Information C-Means,FLICM)畜肉图像分割方法(KFLICM_UG方法和KFLICM_MG方法)。首先利用广义核函数或混合核函数可以有效兼顾学习能力和泛化能力的优势,将图像的每一个像素映射到高维的特征空间,扩大像素有用特征的类间差异,使像素在高维特征空间中拥有更优的线性可聚性;然后结合像素的局部空间和灰度信息,确定其模糊隶属度,在高维的特征空间中依据图像特征对像素进行模糊局部信息C均值聚类,最终实现畜肉图像的分割。大量的实验结果表明,相比现有的模糊C均值(Fuzzy C-Means,FCM)分割方法、KFCM分割方法和FLICM分割方法,本文提出的KFLICM_UG方法和KFLICM_MG方法可以获得更好的分割效果,更低的分割错误率,且具有更强的噪声适应能力和鲁棒性。

Focusing on the lack of a strong adaptive ability to noise of the meat image segmentation method that is based on the traditional kernel fuzzy C-means(KFCM) clustering, two image segmentation methods(KFLICM_UG and KFLICM_MG) were applied to meat samples; these segmentation methods used fuzzy local information C-means clustering(FLICM), which is based on generalized or hybrid kernel function. Firstly, the generalized or hybrid kernel functions were used to strike a good balance between learning and generalization abilities. Each image pixel was mapped onto a high-dimensional feature space, which leads to a larger inter-class difference between the useful features of pixels. Thus, those pixels could be clustered more easily in the high-dimensional feature space. Then, the fuzzy membership of each pixel was determined based on the combination of its local space information with grayscale information. Finally, meat image segmentation was completed via the fuzzy local information C-means clustering according to the image features in the high-dimensional feature space. Considering results of previous studies, compared with the existing FCM(Fuzzy C-Means) segmentation methods such as KFCM and FLICM segmentation methods, the proposed method(KFLICM_UG, KFLICM_MG) can achieve better segmentation results with lower segmentation error rate, stronger adaptiveness, and robustness against noise.