Research On And Implementation Of Cell Image Classification Based On Support Vector Machine And D-S Evidence Theory

Cell image classification is a major issue in the field of image processing and medicine,which is widely applied in the automatic analysis instruments in the medical cell.The conventional approaches applied in the cell image classification include neural networks,decision trees,support vector machines(SVM),etc.As one of the classical method in machine learning——SVM has the characteristics of better robustness and stronger generalization ability for the small samples compared to other machine learning methods.In order to improve the accuracy of cell image recognition by SVM,this thesis analyzes the shape features and texture features of cell image respectively,and then adopted multiple classification SVMs with posterior probability information to recognize cell image,at last these recognition result with posterior probability through different SVMs are fused by D-S evidence theory,which can give the final results of the cell image classification.The main work is as follows:Firstly,to extract the shape features of the cell image better,the method of combining Otsu algorithm and Canny operator edge detection is adopted to detect the edges,and then the background and target of the cell image are separated by combining with the morphological operations.Seven shape features of the cell image were extracted such as area,perimeter,roundness,shape factor,the longest diameter,the shortest diameter and the ratio of the long axis to the short axis,are described and analyzed.Base on the seven features,the multi-class SVMs is applied in the classification experiments on cell images to recognize cell particles.Secondly,to extract the texture features of the cell image better,the gray level co-occurrence matrix and third-order differential invariants method are adopted separately;the latter method is proved to be better.Nine features images of each cell images are derived by the third-order differential operation under the Gaussian smoothing operation with the different coefficient values.The values of sum,mean and variance in the target region of these features images are calculated which are presented as the texture features of each cell image.The multi-class SVMs is also applied in the classification experiments on cell images to recognize cell particles.Finally,considering the unbalance between the number of dimensions of the shape features and the number of dimensions of the texture features,D-S evidence theory is applied to fused the recognitions result by SVMs with the shape and texture features respectively in the information fusion.The twice D-S evidence theory fusion steps are proposed to recognize cell particles.In the first step,the posterior probability values obtained by ν-SVM in the shape features space and the texture features space respectively are transformed into the corresponding values of a proposed basic probability assignment function,and D-S evidence theory is adopted to fuse these values,which could get the reliability assignment value of each kinds of cells,the other corresponding four kinds of cells and uncertain kind.In the second step,D-S evidence theory are applied to fused the results between classes of the first step.After these two fusion steps,the final reliable assignment of each class of cells is obtained.According to the principle of the maximum reliability assignment,the cells are recognized as particle which has the maximum reliable assignment value.