An Algorithm of Image Matching Based on Both Complex Wavelet Energy and SVM
基于复数小波能量特征和支持向量机的图像匹配算法

This algorithm, which is based on both statistical characteristics of complex wavelet energy and SVM, is proposed in order to effectively detect and track targets in image, which may cause changes, such as translation, scaling and rotation. So, the problem of image matching is transformed as that of classification. The transformation of complex wavelet that has properties of scale, shift invariant and directional selectivity effectively extract the statistical characteristics of image, such as mean, standard deviation and skew. The statistical characteristics of sample templates are input into SVM to train support vectors of SVM. Then, those statistical characteristics of any sub-image from original image are input into SVM in order to match target. This is a two-stage algorithm of coarse-to-fine. Firstly, the set of candidates is sifted by SVM. Secondly, a new optimal rule, which is nonlinear distance function, is proposed to decide the optimal matching from the candidate set. Those experimental results show that this algorithm addresses the problem of confidence level, which generally exists in traditional matching methods. This algorithm's performance is superior to those of both learning method of neural network based on RBF and gray-level correlation matching method, which compares with them. Finally, a good matching result is obtained.