Research On Automatic Target Recognition Based On Featrue Fusion

Automatic Target Recognition (ATR) based on information fusion is the technique that integrates complementary information of target images provided by multi-sensors to improve the accuracy and robustness of identification. This is a promising research topic.Much of the current research work focuses on the pixel-level fusion. But pixel-level fusion is to deal with a lot of data which need to be registered in time and space with a high-precision. So it's time-consuming, and can hardly realize real-time process. Decision-level fusion is to integrate the recognition result of each sensor. So it may lose some important information, and get poor performance. Feature-level fusion not only can increase the possibility of extracting features from images, but also can get some useful composite features through integrating existing features. Through feature-level fusion, the dimensionality of data can be greatly reduced by data compression, while sufficient valid information can be maintained. So it can be applied to real-time applications, and get good recognition accuracy. Now, more and more researchers are engaged in the study of feature fusion.ATR based on feature fusion is a technology that fusions target features extracted from two or more sensors to recognize the target.This dissertation focuses on ATR based on feature-level fusion. The research includes feature extraction of target images and fusion methods on feature-level.The main contributions of this dissertation are summarized as follows:1. Based on Normalized Laplacian Locality Preserving Projection (NL-LPP), a novel algorithm called Uncorrelated and Orthogonal Locality Preserving Projection based on Normalized Laplacian (UONL-LPP) is proposed for dealing with the shortcomings of NL-LPP. It shares the same locality preserving character as NL-LPP, but at the same time it requires the basis vectors to be statistically uncorrelated and orthogonal. This makes it not only contain minimum redundancy, but also can reconstruct the original data perfectly.2. As a well-known approach, Linear Discriminant Analysis (LDA) captures the global geometric structure of the training dataset. And Local Preserving Projection (LPP) can get the local geometric structure based on the graph Laplacian. However, structures of real-world data are often complex, and a single characterization (either global or local) may not be sufficient to represent the underlying true structures. In this dissertation, a novel dimensionality reduction algorithm named Global and Local feature fusion Analysis (GLA) is proposed, which integrates both global and local structures. A tuning parameter is employed to balance the tradeoff between global and local structures as different applications. And the recognition rate can be improved effectively. We generalize the GLA approach for extracting nonlinear features via the kernel trick, and thereafter propose the Kernel Global and Local feature fusion Analysis (KGLA).3. Feature fusion based on canonical correlation analysis has been studied in this dissertation. In order to decrease the computational time and avoid the singularity of covariance matrix when canonical correlation analysis is used for feature fusion of images, a fast algorithm was proposed. The algorithm considers an image as the second order tensor. It is based on two-dimensional image matrices rather than vectors. So, variance and covariance can be constructed using the image matrices. After getting the projection matrices by canonical correlation analysis, we can project the image matrices into a space which is the tensor product of two vector spaces. The relationship between the row vectors of the image matrix and that between column vectors can be naturally characterized by the proposed algorithm. The experiments suggest that the proposed algorithm not only speeds up the computational efficiency, but also achieves much higher recognition accuracies.