Research Of Edge Detection Based On Support Vector Machine

Edge is one of the most important characteristics in an image, and edge detection is the base of the image processing. How to detect the edge of the objects exactly and quickly is a hot issue in this area.The quality of an algorithm is mainly reflected in the ability to do the following two aspects: first, it can correctly detect all the edges and does not appear pseudo-edge; second, it can suppress noise as much as possible, i.e. ability of denoising. The main purpose of this paper is, based on the original algorithm, to improve the performance of the filter, achieve the right detection to the edge and raise the quality of the edge detection combined with support vector machine (SVM) theory and technology.At first, some classic edge detection methods and recent fully-fledged ways are introduced, basic concepts and implementation of these techniques are summarized, and their advantages and disadvantages are analyzed with simulation in this paper. Then introduce the multi-scale adaptive filtering technology with a new idea to improve the effectiveness. After this, we introduced SVM, and give the principle and ideas of edge detection based on least squares support vector machine (LS-SVM) with Gaussian radial basis kernel function. By using the proximal support vector machine (PSVM) to fit the surface of the image, a set of new gradient operators and the corresponding second derivative operator are obtained, combined with the non-maximum suppression and dual threshold technology of Canny operator, some better experimental results are attained. Besides, we introduce support vector classification (SVC) into edge detection, through the rational sampling, the obtained edge image is also as a reference to the final output, and this can further improve the test results. The setting of filter parameters, the selection of two thresholds, kernel function and its parameters are discussed respectively in this paper. By comparing the output image and signal-to-noise ratio (SNR), we can conclude that the filter is reasonable, and it is better to white Gaussian noise; with the figure of merit (FOM), a combination of the polynomial kernel function and Gaussian radial basis kernel function is obtained as a new kernel function, and also get the corresponding parameters and the size of the convolution mask. At last, the performance of the proposed algorithm is compared with many other exiting methods, including Sobel and Canny detectors. The experimental results indicated the superiority of the proposed edge detector. Experiments show that the method of this paper is feasible. Also it expands the application of SVM theory and provides a new way of thinking for edge detection.