Study Of Target Recognition Based On Mutisensor Image Fusion

Target recognition based on multisensor image fusion is an important study problem and has great value in application in the field of image reconnaissance. This thesis carries out the study of theory analyzing and data simulating on image fusion processing, image feature extraction and choice, target feature template database conformation and image target recognition, etc. The study fruits we gained show as follows:1. PCA-based wavelet coefficient fusion method has been presented. And it can adaptively set low frequency fusion coefficient in wavelet domain and solves effectively the problem of low frequency fusion rule coefficient setting about images that have local contrast reversal, common features and complementary features.2. These wavelet image fusion based on high coefficient region 'energy' for visual and infrared images, the one on high coefficient local contrast for visual multi-focus images and the one on high coefficient grads for visual multi-focus images are presented. And the efficacy of them is validated by experiments.3. A multi-feature extraction method combining separability criterion with K-L translation is advanced and can be used to choice, decorrelation and reduce dimension, so as to reduce computing greatly.4. A hierarchical subtractive fuzzy clustering algorithm is putted forward in this thesis. The method needn't know the number of target clusters to obtain target clusternumber by splitting automatically and to cluster. Moreover, it has higher computing efficiency comparing with traditional fuzzy c-means.5. Fuzzy covariance-based adaptive clustering neural networks (FCACNN) is presented. It combines the competitive learning of the neural networks and the distance measurement of the fuzzy covariance clustering algorithm to cluster steadily by combining the resemble clusters iteratively.6. For image classification and recognition, the approaches to set up sample feature database using self-organizing feature map neural networks (SOM) is presented in this thesis.7. Based on maximal-minimal membership rule, k-nearest neighbors image target recognition with adaptive threshold is advanced. It cannot only recognition the image target learned, but also can recognition unknown image target.