Research On The Joint Classification Based On Support Vector Machine And K-nearest Neighbor

With the rapid development of science and technique, the active sonar applications are widespread, recognition technique on underwater targets as a crucial sonar technology is also developed. However, compared with other several sonar techniques, the development of active target classification still appears slowly. There are many factors which restrict its development. So how to extract the feature of object effectively and design good classifier is a key problem in the underwater targets recognition.The Support Vector Machine which is based on Statistical Learning Theory shows strong adaptability, excellent study ability and generalization function in solving small sample learning problems. It is so difficult to get samples from the underwater target echo that the number is limited. The Support Vector Machine is thought to be the most appropriate classifier for the underwater target recognition. However, the ability of a single classifier is often limited, it often make mistakes nearby classify hyperplane, so we must make use of the samples information nearby SVM hyperplane. We do the research on the basic theory and algorithm of SVM, Improved KNN-SVM that combined Support Vector Machine (SVM) with K-nearest Neighbor (KNN) is presented to improve the accuracy of targets recognition nearby SVM hyperplane. K-nearest Neighbor classifier can make all support vectors as its representative point. The graduate thesis does the further research and adopts the joint classifier based on Support Vector Machine and K-nearest Neighbor as the classifier in the underwater target recognition.Feature extraction approaches based on Wavelet Packet Decomposition,Wigner-Ville Distribution and Fractional Fourier Transform are discussed, and the ways to form the feature are proposed. The paper uses principal component analysis as reduced-dimension method. Extracting the features from two kinds of suspended and buried target echoes and classifying them are accomplished. The experimental results demonstrate that the features are feasible. SVM-KNN is more robust than the traditional SVM and K-nearest Neighbors.