Intelligent Recognition Key Technology For Corporeal Components In Humoral Cell Image

The automatic recognition and analysis of iatrical microscope image is very active in biomedical engineering domain. It can improve the assay's efficency and provide more scientific results by achieving corporeal components' automatic recognition and analysis in humoral(urine)cell image. Intelligent recognition key technology for corporeal components in microscopic image was researched in this dissertation, which mainly involves cell image preprocessing, image segmentation, shape feature extraction, texture and color feature extraction and the target recognition.Image preprocessing includes two ways such as filtering noise and enhancing the edge.The dissertation firstly discussed the factors of lowering quality and analyzed the noise model. Secondly,three algorithms were proposed, which are the anisotropic filter algorithm based on the gradient and adaptive median filter algorithm union information entropy and edge enhancement algorithm based on the neighborhood region contrast. The image background noise and the pulse noise have been seperately filtered by the ameliorative anisotropy fliter and the adaptive median filter, at the same time the detail information were reserved. The edge of image was enhanced by calculating neighborhood contrast and designing appropriate nonlinear function,.The accurate and correct segmentation is the foundation of cell image recognition.The thesis firstly discussed many kinds of auto-adapted segmentation technique model and the multi-level segmentation algorithm based on evaluation and designed the algorithm of cell image synthetical segmentation,which is multi-information and multi-level and multi-method and achieved accurate and correct segmentation. In order to achieve multi-level segmentation,the thises proposed two new methods,one of which is the multi-level double value iteration segmentation algorithm based on BP evaluation and the other is multi-level adaptive edge detecting algorithm based on improved canny operator. To realize multi-imformation segmentation,the thesis put forward two methods,one of which is a fast and improved 2D threshold segmentation algorithm grounded on gray value and the average value of neighbour region, and the other is a mean shift segmentation algorithm, which synthesized the information of space, gray value and color. In order to overcome rupture and false connection of the edge the morphology multi-structure units edge joint algorithm has been used after each segmentation.Shape feature is one kind of important information for distinguishing corporeal components in cell image. The thesis firstly discussed the primary influencing factors and established each corporeal component's shape semantic model.Region chord distribution shape descriptor and fourier conversion coefficient descriptor were defined. One kind of shape description method based on polygonal approximation and curvature was proposed. A rapid calculation method for moment feature based on the boundary point was proposed. Ultimately, 35 kinds of robust shape feature in distinguishing the shape differences were established by above methods.Texture and color feature is another kind of important information. the semantic description model was established by observeing corporeal component's atlas. The synthetical commensal matrix texture description method was elaborated in this dissertation. Besides, some methods were proposed including a description method which combined texture scores with Zernike moment, a method which extracted the texture feature by multi-scale high frequency energy ratio and the orientation of texture transformation based on wavelet multi-scale analysis and a method used in color feature extraction based on probability sliding window. Finally, 32 kinds of texture and color feature describing each component's interior difference were established.Cell pattern recognition was realized after extracting all kinds of feature. The dissertation discussed the basic framework of cell pattern recognition and achieved it by BP neural network and ID3 decision tree method. Besides,the thesis compared the capability of these two methods. and achieved cell pattern's recognition by BP.On the other hand a kind of normalization algorithm based on the probability distribution equalization was proposed, which quickened the studying convergence time and improved the recognition ratio.