Research On Automatic Target Recognition In High Resolution SAR Images

hi this dissertation we present an automatic target recognition system with application to high-resolution synthetic aperture radar imagery. A multistage SAR ATR framework is proposed. The system is composed of three main stages, the detection, the discrimination and the classification. The achievement of this research will improve the military reconnaissance ability of our country.The high-resolution SAR image-modeling problem is first studied, and then a high-speed target detection algorithm based on CFAR technique and target variance character is proposed. This CFAR detector is composed of two processes, the horizontal CFAR and the vertical CFAR. The superposition of adjacent reference windows and the distribution character of the image are used to speedup the parameter estimation. The variance character of target is utilized to reduce the effect of bright natural clutter, such as trees, so as to reduce the false alarm rate.Four target discriminators, the Area Discriminator (AR), the Peak Power Ratio discriminator (PPR), the Vector Quantization discriminator (VQ), and the Extended Fractal feature discriminator (EF) are proposed. The AR discriminator is simple and effective. The PPR discriminator is also easy to compute, and is effective in eliminating the false alarms caused by natural clutter, such as trees. The VQ and EF discriminators are a bit complex, but they are based on the silhouette of target, and they can eliminate the false alarms caused by the bright clutter of man-made scenery, such as buildings. A discrimination algorithm integrated these discriminators is proposed. This algorithm can discard most of the false alarms while maintain a high detection rate. Before discrimination, we present two ROI segmentation algorithms to extract the target region, one is an ICM segmentation algorithm based on MRF image model, and the other is based on CFAR technique.We propose an azimuth estimation algorithm based on the near radar edge of target. A linear fit is performed on the major edge of the range direction or the azimuth direction according to the characteristic of the near radar edge. The images segmented by the two segmentation algorithms are used to test the azimuth estimation algorithm, and the result is good. The algorithm is not only very efficient, but also robust to certain segmentation error.The azimuth angle and depression angle variance of the peak feature are first studied. Then we propose a target classification algorithm based on peak features. The matching problem is formulated as a non-linear objective function, which incorporates both the location and magnitude of the features. The minimum of this function is found using a combination of deterministic annealing and softassign. The classification ability of three peak feature classes is compared, and the problem of synthesizing multiple feature classesis studied. We also study the classification performance of occluded targets. Experimental results using simulated occluded target images show that excellent recognition results can be achieved at very high occlusion.