Study On Segmentation And Recognition Method Of Red And White Cell In Urine Visible Components Image

It is one of the important bases for the diagnosis of urinary and renal diseases that judging the type of specific cells and their quantities in the urine forming component.This study designed a method for segmentation and classification of red blood cells and white blood cells in the urine image based on low-dimensional feature conditions,which not only ensures the specific cells extracted from the background under low magnification(acquired under 20-fold eyepiece system),but also improved the recognition accuracy to a certain degree.The main work is described as follows:1.We designed an image denoising and enhancement method combining Gaussian Mixture Model and distance matrix.Firstly,selected Sum Average derived from gray level co-occurrence matrix as texture features.Gauss mixture model is established for urine visible components image.The image is enhanced by the interaction of neighborhood pixels to reduce the influence of noise and weak edge phenomenon on segmentation.2.A method of urine adhesion cell segmentation based on Markov Random Field(MRF)was designed.Considering the obvious difference between background and foreground texture,the idea of texture classification is used to achieve the purpose of segmentation.Similar weights are introduced into potential function.Markov model is used for texture segmentation of feature image.The interaction between pixel marker information and texture information in potential group is considered.The area ratio method is used to find the candidate set of concave points of adhesion cells,and a concave clustering method is proposed to find the best pair of concave points for segmentation,so as to complete the segmentation of adhesion cells.3.A topological feature vector extraction method is proposed.The number of regions,the gradient space variance,the maximum area ratio,the perimeter and the area are used to form the feature vector.The relationship between features is found to establish feature network for the classifier training.4.The experiments of HMM,SVM,SVM-HMM,low-dimensional feature combination and common feature combination recognition are designed.Experiments on 300 samples of erythrocyte showed that the recognition effect of SVM-HMM based on improved low-dimensional eigenvector was better than HMM and SVM.The recognition rate of erythrocyte and white blood cell was 95.7% and 97.8%,respectively.The standard deviation of the recognition rate of grouping test was small and the recognition effect was stable.It is hopeful to be further applied in clinical microscopy.