Recognition Of Spermatogonium Image Using Support Vector Machine

Recently, with the development of information technology, digital image processing andrecognition technology have been developing rapidly, and the theory on further researchgrows constantly. It has been applied in various fields, for example, military, publicsecurity, aerospace and satellite remote sensing image recognition, identification ofmonitoring system, mobile internet and biomedical image recognition et al. At present, thecell recognition is manually done by experienced doctors according to biologicalparameters of cells, it is obviously a boring and time-consuming job. Meanwhile the effectof recognition may be easily influenced by the doctor’s experience. Doctors are liable tomistake in recognition due to continuous work or inattention. Particularly, previousresearch has been focused on the intact cell recognition, and there is little research on cellslice image recognition. However, cell slice image recognition has great significance fordiagnosis and3D reconstruction. Therefore, in this article, slice image recognition ofspermatogonia is researched based on digital image processing, image feature extractionand machine learning.There are three main tasks for spermatogonia image recognition: image preprocessing,image feature extraction and image recognition.Firstly, the paper introduces the common methods of preprocessing, generating thebinary images of the target cells by digital image processing methods and mathematicalmorphology. Then the concept of image feature and general image features are described.Due to the characteristics of the images, the magnitudes of Zernike moments is used torepresent the image features. Besides, the methods of corrected phases of Zernikemoments are also introduced to form another type of images features combining with themagnitudes of Zernike moments.Finally, Support Vector Machine (SVM) is explored to train and recognize all the images. During the training step, different size of training set are setup, and RBF kernelfunction is used by introducing penalty factor C. In order to verify the reliability of theexperimental results, the same scheme is imposed on another image database: MNISTdatabase. Identification of spermatogia images based on Euclidean distance is alsocalculated to compare with the proposed methods.Experimental results show that the recognition based on the magnitudes of Zernikemoments and SVM has high efficiency and accuracy and outperforms the traditionalmethod using Euclidean distance.