Edge Detection Based On Least Squares Support Vector Machine With Mixtures Of Kernels

In image processing, as a basic characteristic, the edge of the image, which is widely used in the recognition, segmentation, intensification and compress of the image, is often applied to high-level domain.There are many kinds of ways to detect the edge. The method, which is got the longest research, gets the edge according to the variety of the pixel gray. The main techniques are Robert, Laplace, Sobel, Canny and LOG algorithm. But in practical application this method subject to certain limitations because it is sensitive to noise. In recent decades, many scholars around the edge detection algorithm for a lot of problems, such as the wavelet method, fuzzy reasoning, neural networks, surface fitting methods. In these algorithms, surface fitting method is the relatively good in recent years, because of the strong anti-noise detection and the high accuracy, this algorithm is a more mature developed method now. Support Vector Machine(SVM) method is a surface fitting algorithm in a way.SVM is a new pattern recognition technology that is established on Statistical Learning Theory. It can solve small-sample learning problems better by using Experiential Risk Minimization. Moreover, this theory can change the problem in non-linearity space to that in the linearity space in order to reduce the algorithm complexity by using the kernel function idea.On the basis of studying on least-squares support vector machines (LS-SVM) of the image edge detection technology, Proposed a new method, which is based on mixed Kernel LS-SVM image edge detection. This algorithm based on the combination result of gradient and zero crossings is presented which the image intensity of neighborhood region of pixel is well estimated by LS-SVM with mixtures of kernels and the gradient operator and zero crossing operator are obtained by LS-SVM based on mixtures kernel function. The result of experiment shows that the LS-SVM performance by using mixtures of kernels is much better than that using polynomial and Gaussian kernel function when the SNR is lower.