Recognition of Remote Sensing Target Based on Support Vector Machine
Zhang Yanning, Zheng Jiangbin, Liao Yi, Zhao Rongchun (Department of Computer Science and Engineering, Northwestern Polytechnical University, Xi′an 710072)
The recognition results of image targets by existing methods (such as neural network and Euclid distance methods) are not satisfactory for shaded image or 3-D rotation image. We present an improved SVM (support vector machine) method for recognition of such images. First, remote sensing images are processed into binary images. Then, the binary images are nomalized according to the kernel function used in SVM (where the kernel function is Gaussian, the range of normalization is 0 to 0.02), and the normalized image targets are divided into two sets, one is for training, another is for testing. After that, SVM is trained by the training set. Finally, the trained SVM is used to test the testing set. We find that the normalization process is crucial for the application of SVM. Our main contribution is in the formula for normalization of targets as given in Eq. (7), where C is a constant which is determined by the kernel function and experiments. Table 1 shows the effect of constant C on the correct recognition rate. Comparisons as given in Table 2 show that the correct recognition rate of the improved SVM method is 96%, while that of neural network method and Euclid distance method are 90% and 84% respectively.