Application Research On Multi-Sorts Targets Recognition Based On Improved SVM

Technique of underwater targets recognition and classification is paid much attention by researchers as its specially important status in national defensive field. Combining latest machine learning tool—Support Vector Machine (SVM) in statical learning theory, the research on underwater targets recognition and classification with SVM technique is processed in the thesis. Basic theory, forming methods of SVM is introduced detailedly, and design of rapid algorithm is developed, the method of multi-patterns classification is constructed, which offers stable theoretical foundation for the underwater targets recognition and classification. Based on these, features extraction to the radiated noise of practical underwater targets is completed, while classification experimentation with new SVM algorithm is finished, satisfactory results have been got. The main contents and innovation are summarized as follows:1. Basic features of radiated noise of underwater targets is analyzed, including source style, time domain features, and spectrum features, especially "three not" features, that is transfering channel and receiving system are nonlinear, noise signal is not Gaussian and not stable.2. Based on high-order statics, characteristics of double spectrum and 1(1/2)-dimension spectrum, double cepstrum and its lower dimension spectrum are studied. Structure and characteristics of 1(1/2) -dimension spectrum and double cepstrum's lower dimension spectrum of underwater targets are researched, finally, 50 dimension characteristics of 1(1/2) -dimension spectrum sub-band energy and 25 dimension characteristics of double cepstrum's lower dimension spectrum time-domain energy in three kind underwater targets radiated noise are constructed.3. Based on wavelet transforming theory, feature diversity of three kind underwater targets radiated noise in 5-level scale is studied, the attenuating factor of scale-energy of radiated noise is difined, scale-energy characteristic of underwater targets radiated noise is finally conceived.4. According to the practice of large size training samples, rapid training algorithm—SMO algorithm which performs excellently in training time and memory using is studied. Based on this, Cache-SMO algorithm is produced while using proper cache-memory strategy.