Target Recognition Of SAR Images Based On Neural Network Ensemble

In recent years, neural networks ensemble has been widely used in image processing and the effect is remarkable, especially in extracting useful information rapidly from large-scale SAR data. On the basis of neural networks ensemble learning combining with SAR image processing, this thesis concentrates on SAR image target recognition based on neural networks ensemble learning. Specific works and innovations are as follows:Firstly, considering various disadvantages of feature extraction in the original image domain, such as high feature dimensions, large computational complexity and weak ability to represent classification, feature compression and feature selection are applied to improve feature extraction. The improvement of feature compression: due to the fact that the low-frequency sub-image fully represents the target information, two-dimensional discrete wavelet transformation is used between preprocessing and feature extraction, so that certainty target information is separated from random background clutter and noise. Then, the feature of the low-frequency sub-image is extracted by Bi-2DPCA, which can effectively remove redundant information between rows and columns, but also make low dimensional target feature good coherence within class and differences between classes. The improvement of feature selection: K-NN algorithm is used to achieve adaptive feature selection of training samples and classifier training. After the two improvements above, the optimal target feature not only has lower dimension but also better recognition performance, which has important significance to promote intelligent SAR target recognition based on neural network.Secondly, by comparison with performance of BP and PNN based on experimental simulation, PNN with more real-time and more stability is selected as the integrated individual network. In order to solve the problem that the PNN structure is complex and training algorithm is not adaptive, adaptive training algorithm based on neighbor subspace is proposed. Simulation results show that the efficiency and recognition accuracy of the improved adaptive PNN are better.Lastly, in order to overcome the shortcomings that generalization capacity of adaptive PNN is weak and recognition result of adaptive PNN is unstable, combining with the idea of ensemble learning, a new recognition method is presented that a group of individual adaptive PNN classifiers are combined by bagging algorithm, and experiment results show that bagging-PNN can effectively improve recognition efficiency, recognition accuracy, generalization ability and recognition stability. Based on MSTAR SAR database, comprehensive experiment results prove that the proposed algorithm is effective and practical to SAR target recognition.