Classifier Design And Subspace Learning For Radar Emitter Recognition

Radar emitter recognition plays an important role in the ECM. With the rapid development of the radar technology and the utility of new types of radar, exploring more effective feature extraction methods is urgent from the point view of signal processing. From the perspective of pattern recognition, it is also essential to concentrate on the classifier design and feature optimization. Based on the two effective features of radar signals, this thesis deals with classifier design, as well as subspace learning based dimensionality reduction technology, for radar emitter recognition.For feature extraction, zero-slice of the cyclic spectrum and concatenation of the slices in ambiguity function based methods are studied respectively. The experimental results on the real radar data show the validity of the two methods, which provide reliable features for subsequent classification tasks.For classifier design, six types of classifiers which can produce posterior probability are introduced. Then classifier combination can be conveniently realized via proper fusion strategy. To further reject the out-of-database targets, generalized confidence is calculated by the posterior probability, and then thresholds are empirically chosen to implement the rejection operation. Finally, several evaluation indices are given. The experimental results demonstrate the advantage of classifier combination, as well as the feasibility of generalized confidence based rejection methods.For subspace based dimensionality reduction, both linear and kernel discriminant subspace learning are investigated in depth. Therein, kernel-based methods are the extensions of corresponding linear methods via some specific kernel trick. These classical algorithms are summarized into single-subspace learning and multi-subspace learning, which are applied to image recognition and radar emitter recognition. The experimental results indicate that multi-subspace learning algorithms take into account the complementarity of different subspaces and are more robust than single-subspace based methods, no matter linear or kernel cases are concerned. Due to data dependence, kernel extension not always guarantees higher accuracy, so appropriate method needs to be selected according to the actual data distribution and engineering requirements.