| It is a very hard problem to accurately, quickly and adaptively implement image segmentation in the field of image analysis and pattern recognition. The results of segmentation will directly affect the successive processing. Since the multispectral microscope cell image can provide enough information, we present a method in which we classify the image at first and then segment it by using the image pixel as the recognition object and gray values of the pixel's wave bands as feature. Based on Statistical Learning Theory (SLT) and Support Vector Machines (SVM), we study and implement a segmentation algorithm of the multispectral microscope cell image in this paper. The experimental results prove that the algorithm is of high precision, quick and adaptive.In this paper we introduce two models-Radial Basis Function NeuralNetworks (RBFNN) and Support Vector Machines (SVM). The model's principle, characteristics as long as the learning and classifying algorithms are discussed in details. We implement the classification of multispectral microscope cell image by using the two classifiers, in which each wave band's gray value of image pixels are used as feature. Following that we compare the two classification algorithms on the basis of the experimental results. This is the main work of this paper.Since there is a vast amount of data volume in the multispectral image, and our research result will be a part of an application system, we must preprocess the multispectral microscope cell image before classifying in order to reduce the computation time. Based on the study of the image spectral characteristic, we propose the image segmentation and preprocess methods of microscope cell image. In preprocessing, we transfer the RGB space of the multispectral microscope cell image into the HSI space, and remove the background and parts of red cell of the multispectral microscope cell image using the threshold segmentation method on the intensity weight and hue weight. After the preprocessing the remained pixels of the image are classified by RBFNN classifier and SVM classifier. The preprocessing can increase the computation speed to a large extent.The pre-segmented image has three intensity value (indicating background, white cell kernel and white cell plasm respectively), a few of small grains and holls in the image can't be avoided. In order to increase the accuracy of white cell segmentation and recognition in diseases diagnosis, we must process the three-value image more precisely. On basis of introducing some region growing criterions and methods of region growing algorithm, we adopt the criterion based on gray value difference (gray value difference is zero) to remove the small areas or small holls whose size are less than a threhold. By this means a good segmentation, which establishs the basis ofsuccessive feature extraction and recognition diagnosis, can be obtained.We provide 220 multispectral microscope marrow cell images as experiment samples. The experimental results indicate that the segmentation accurate rates of RBFNN and SVM are 85% and 87.7% respectively. The SVM approach has short training and classifying time, so SVM is more fit to clinic diagnosis. Our methods are also applyed to other type multispectral microscope cell images and can gain very good results. It sufficiently proves that our approach is efficient on the segmentation of multispectral cell images. We find out that SVM has better performance in classification and recognition due to the less training and classifying time and higher accurate segmentation rate. We also compare RBFNN and SVM on theories to some extent in the end.The experimental results prove that the methods presented in this paper can segment the multispectral cell images more efficiently and precisely than the traditional segmentation algorithms. As a part of the "the Research and Development of the Imaging spectrometers", which is the China Tenth Five-year Brainstorm project, our research result has provided a good foundation for the project.
|
PDF Full Text Request