Research On Multi-feature Fusion Classification And Recognition Of Cell Image

Quantitative analysis and automatic recognition of cell pathological images by computer technology has important theoretical significance and application value.Based on the knowledge of image processing and pattern recognition,this thesis proposes a cell image classification and recognition method which combines multi-feature fusion and model fusion.It focuses on cell nucleus segmentation,feature parameter extraction,feature-level fusion and decision-level fusion to realize the recognition and classification of cell pathological images.The main research work completed is as follows:1.Improvement of ACWE model segmentation methodThe improvement of ACWE model segmentation method in this thesis is mainly reflected in two aspects: one is the determination of the initial curve of ACWE model.Firstly,the cell image is enhanced by Gauss filtering and logarithmic LOG transform.Then,it is transferred to Lab color space to extract the B-component space image.Finally,the B-component space image is preliminarily segmented by marker-based watershed algorithm,and the outer rectangle of the segmented contour is obtained,which is used as the initial curve of ACWE model.Second,aiming at the drawbacks of large amount of calculation and poor real-time performance caused by artificially setting area coefficient in ACWE model segmentation,an adaptive area coefficient calculation method is proposed.Compared with the traditional ACWE model segmentation method,the improved ACWE model segmentation method not only reduces the number of iterations,improves the efficiency of cell nucleus segmentation,but also achieves the accurate segmentation of cell nucleus,which is more conducive to the extraction of morphological features of cell nucleus in the later stage.2.Feature parameter extraction of cell imagesFirstly,a texture feature extraction method is proposed by combining LBP operator with Gabor transform,and GLCM feature is taken as the global feature of cell image.Secondly,morphological features,such as area,perimeter,aspect ratio,roundness,rectangularity and nucleocytoplasmic ratio,were extracted from the segmented nuclear region as local features of the cell image.Finally,the three features are cascaded to get the feature set,and the feature set is normalized.3.Feature level fusion algorithm of PCA-RFECVThe normalized feature set is fused by principal component analysis(PCA)to get the reduced feature set.Then the cross recursive feature elimination algorithm(RFECV)is used to select the second feature set after dimensionality reduction,and the most representative feature is selected as the description feature of the cell image.4.The decision-level fusion method based on Stacking model fusionThe Stacking model in this thesis is divided into two layers.The model of the first layer is called the first-level model,including random forest,decision tree,k-neighborhood algorithm and gaussian bayesian classifier.The model of the second layer is called the second-level model,and the support vector machine(SVM)and gaussian bayesian classifier are used as the second-level models in the bi-classification experiment and tri-classification experiment respectively to complete the classification and recognition of cell images.Stacking model fusion method improves the classification recognition rate to a certain extent.5.Experimental simulation validationThe cell image segmentation,feature parameter extraction,feature set fusion and decision level fusion algorithm are simulated numerically.By comparing with other methods,the validity of the segmentation method,feature set fusion method and decision-level fusion method is verified.Figure [38] table [9] reference [55]...