利用单类支持向量机分割血细胞图像

Segmentation of Blood Images Using One-Class Support Vector Machine

为了提高白细胞自动识别算法的性能,提出了基于均值移动和单类支持向量机的血细胞图像分割新方法.该方法的原理是将图像中颜色相对稳定的背景和红细胞部分像素作为正训练样本,将颜色复杂多样的白细胞像素作为异常数据检测.均值移动过程用来在红、绿、兰(RGB)颜色空间寻找正训练样本集,通过均匀抽样和颜色量化措施,实现单类支持向量机(SVM)在线实时训练,最终图像像素经过单类SVM分类来实现分割.实验表明,新方法对涂片制备和光照变化导致的图像颜色改变有很好的适应性,图像分割精度优于常用流域算法,而耗时只是后者的1/4.

A new method for segmentation of blood micrograph images is proposed which combines the mean-shift algorithm with one-class support vector machine (SVM) to extract white blood cells. The one-class SVM model is trained by the positive samples made up of some background and red blood cell pixels, then the white blood cell pixels are detected as outlier data by the model. The mean-shift procedures are used to find the clustering modes of positive samples in RGB space. By means of uniform sampling and color quantization, the size of training set can be limited within 1000, so the training can be completed near to real time. The experimental results on 200 blood images demonstrate the higher segmentation accuracy and efficiency of the new method than the traditional watershed algorithm, and it takes only a quarter of the later's operating time. It brings robust performance to cope with the change of color results from varied preparation and illumination in image acquisition.