Classification Of Urinary Sediment Cell Images Based On CNN,SVM And D-S Evidence Theory

Cell image classification,which is widely used in automatic cell analyzer,is an important subject in the field of image and medicine.In the field of machine learning,cell images are classified by means of convolution neural network,decision tree,support vector machine,bias classification and so on.Compared with other classical learning methods,convolution neural network doesn't need to extract features artificially as the most popular algorithm.So,its recognition accuracy is very high.As the most classical classification algorithm before the second rise of convolution neural network,support vector machine has better robustness and small sample generalization ability than other machine learning methods.In order to improve the accuracy of cell image recognition,this paper constructs a convolution neural network to recognize the cell images at first.Then support vector machine is used to classify them.At last,the probability value of the output of convolution neural network and of support vector machine conversion are fused by D-S evidence theory to give the final result of cell classification.The main tasks include the following contents.Firstly,in order to make the image with different sizes in data set,this paper uses the difference algorithm,fast Fourier transform and an image cutting method improved by the author to unify the size of the images.By referring to the framework design of classical convolution network and then combining with the characteristics of this data set,it can determine the layers of CNN and full connections,the sizes and numbers of convolution kernels,activation function,training mode and other parameters.In addition,it can also set up convolution network for this data set.The training data set determines the algorithm for image preprocessing according to the classification accuracy and whether the image processing method has good mobility.Secondly,the feature extraction of cell images is performed through the GLCM and Gabor filter and the multi-value classification problem of SVM are decomposed into binary-classification problem.The extracted features are applied to train support vector machine.Compared to the results of classification,the model of support vector machine can be determined by error rate and stability of classification.(For different test sets,the error rate of classification won't have a great change.)For data sets with different combinations,applying support vector machine to classify them and then counting the output results according to voting method can obtain the final classification results.Finally,absolute value,which is the output value of classification function of support vector machine,is used to represent the distance from sample point to the hyperplane.Assuminng that the distance meets standard normal distribution,and then converts the absolute value of the output value to the probability value of the classification.According to the principle of the voting method,the classification probability values of binary classification SVM are converted to multi-value.The D-S information fusion of multi-classification probability values of SVM and output probability values of CNN will obtain final reliability distribution for each type of cell.Through the principle maximum reliability distribution,cells can be distinguished that whether they belong to the types of the maximum reliability distribution or not.