Application Of Kernel Theories On Image Segmentation

In this paper, the characteristic and applicable situation to three common kernels (polynomial function, radial basis function and sigmoid function) has been analyzed. An algorithm of fingerprint image segmentation based on Support Vector Machine (SVM) and noise elimination operations has been proposed. Several simulation experiments have been done on Matlab 7.0 and FVC2002 fingerprint databases.Image segmentation was a key technology to image processing and analysis. Its primary task is to separate targets from background or different areas to be distinguished out. But influenced by noise and segmentation algorithms, some useful information was lost in those common image segmentation algorithms and leading to a distortion. Now classified technologies based on machine learning theories become more and more popular in image segmentation. SVM is an attractor in these technologies.Support vectors with great distinguishing capability can be selected automatically with SVM and a classifier was constructed which maximizes the distance between two classes. Effect of the image segmentation depends on the selection of kernel function. Usually three kinds of kernel function are applied to SVM, which are polynomial function, radial basis function and sigmoid function. Characteristic comparison of three kernel function is presented by two simulated experiments. The rules of kernel function selection in different recognition matter are presented. Finally experiments of segmentation to foreground/background of fingerprint images were tested. Here fingerprint image, gradient coherence and gray mean and variance was adopted as input feature vector. SVM was introduced as classifying and judging technology. Intelligent noise elimination operations are also used to optimize the segmentation in the end. The experimental results demonstrate the efficiency and accuracy of these methods.