基于非参数变换的尿沉渣细胞图像识别方法

Urine sediment cell image recognition method based on non-parametric transform

首先提出一种改进的局部秩变换(D-LRT)方法,通过计算图像像素的局部秩进行图像变换,根据阈值选取实现低对比度和局部模糊的尿沉渣细胞图像分割。然后提出一种局部直方图统计(LHS)方法提取细胞图像平移、旋转、光照不变特征。该方法通过高斯模糊获取不同尺度下的细胞高斯模糊图像,并利用RIUP-LBP算子获取细胞高斯模糊图像的RIUP-LBP特征图谱,基于距离变换采取由内而外的方式,分层统计细胞高斯模糊图像及其RIUP-LBP特征图谱的直方图,将所有直方图串联作为细胞图像的LHS局部特征。还同时将LHS特征结合图像的几何特征、Harris角点及灰度共生矩阵特征,从局部和全局角度构造尿沉渣细胞图像的特征向量。最后采用支持向量机(SVM)对7类典型尿沉渣细胞图像进行分类。实验结果表明:本文提出的D-LRT方法在低对比度和局部模糊的细胞图像分割中完整度明显提高,提出的LHS方法可以有效地提取7类尿沉渣细胞图像特征,7类细胞图像识别的平均准确率可达到93.0%,平均召回率可达93.2%。

This paper presents a developed local rank transform（ D-LRT） method,in which the image transform is conducted through calculating the local rank of every pixels in the image. Through a threshold selection step,the segmentation of low-contrast and localfuzzy urine sediment cell images is accomplished. Subsequently,a local histogram statistics（ LHS） method is proposed to extract the cell image characteristics that are invariant to shift,rotation and illumination. The LHS method performs Gaussian blur on cell image first to obtain the cell Gaussian blur image,and the RIUP-LBP operator is employed on the blur image later to obtain the rotation invariance with uniform local binary pattern（ RIUP-LBP） feature maps of the cell Gaussian blur image. Based on the distance transformation,the histograms of the Gaussian blur image and the RIUP-LBP feature map are counted hierarchically. At last,all the histograms are cascaded to form the LHS feature of the cell image. The image geometrical characteristic,Harris corner and grey-level co-occurrence matrix characteristic are also created and integrated with the LHS feature to construct the feature vectors of the urine sediment cell images in local and global perspectives. Finally,the support vector machine（ SVM） was used to classify seven kinds of typical urine sediment cell images. The experiment results demonstrate that when dealing with the low-contrast and local-fuzzy cell images using D-LRT,the segmentation accuracy is increased obviously; the proposed LHS method can extract the features of the urine sediment cell images effectively,the average precision rate of 93. 0% and average recall rate of 93. 2% are achieved for the seven kinds of cell images using the proposed method.