| In the field of remote sensing telemetry, image acquisition and its feature extraction have great effect on target recognition. With the advancement of remote sensing and telemetry technology, Synthetic Aperture Radar (SAR) has unpredictable prospects for military targets, ground civil facilities and natural resource exploration.The speckle noise and complex noise background information can directly or indirectly affect the image denoising, segmentation and feature extraction of synthetic aperture radar. With the development of the imaging technology and its high-resolution images,the images can more vividly express the target’s multi-feature information, such as the texture, structure, morphology and scattering. So its multi-feature can be extracted and integrated to improve the accuracy in object recognition. Therefore, this thesis studies on the geometric feature extraction of SAR 3D image, and carries on its denoising segmentation, and focuses on improving the geometric feature extraction method.Firstly, the thesis does the research on SAR image and the principle of generation background of the noise’s in common condition. And the SAR image denoising algorithms are explored and analysized. Meanwhile, the time performance factor and several common evaluating algorithms like ENL, EPI, are used to specifically quantify the denoising performance of the algorithms in the actual situation. And the simulation verifies that algorithms above are effective.Secondly, aiming at the unpredictable problem of the environment when the SAR image forming, the algorithm of image segmentation based on threshold, geometric form, neural network and collection method is analyzed respectively. And the advantages and disadvantages of above methods are evaluated by simulation experiments, and then the vehicle image segmentation method based on the Weibull distribution CA-CFAR detector is proposed.Finally, the feature extraction algorithm of point and region of synthetic aperture radar image vehicle are studied systematically, and the fusion criteria for multiple images is improved, and its geometric features are extracted according to the shadow of the vehicle. In the meantime, the CFAR detection method, which using cosine Fourier moment and Weibull distribution, is proposed to extract the moments and peak features respectively. The cascade combinatorial classifier is applied to model in target recognition, and it realizes the target recognition when the characteristic dimension is high or the posture changes. Basing on the MSTAR database, the simulation and analysis of SAR images are carried out, and the time performance and algorithm performance of both the image processing methods and steps are analyzed. The reseult is as follows: If the amount of the image input is three on the process of image fusion,the length, width and height of the vehicle are improved obviously, and the result is pretty good. The relative error is basically kept at below 15%, the recognition rates of T-72, BMP2, BTR70 are over 94%, especially the rate of BTR70 is nearly 99%. |
