A Study Of Automatic Target Recognition In Synthetic Aperture Radar Imagery

Synthetic Aperture Radar(SAR) can provide all-weather terms high resolution topography imagery in detail, and also capable of penetrating through observation view port ,which is significant to reconnaissance task. Because of these advantages, SAR is widely applied in many fields such as geological exploration, ground and military surveillance, etc. And also make SAR image processing and research become a hot topic in signal processing domain .This paper mainly focused on Automatic Target Recognition (ATR ) in SAR imagery. In the field' s of SAR ATR, the leading techniques include template matching, pattern recognition and model based technique which is developed recent years. The template matching use sample as the template, calculating cross-correlation coefficient and get the recognition result. The last two techniques build feature library through extracting features from samples, and use these features to match the sensed image and achieve the goal of classification. Template matching needs corresponding relationship between pixels in the sensed image and the pattern, which makes false matching in the blurred image.Several feature extraction methods used frequently are discussed in the paper, co-occurrence matrix based on texture feature and the invariant moment feature based on shape are analyzed in detail, thesefeatures are not appropriate while concerning SAR imagery. We suggestedr Hausdorff Distance(HD) image as the feature which based onedge-extraction, HD is capable of overcoming translation, rotation and scaling, which is a reasonable feature in SAR imagery recognition.In this paper, a three-stage ATR system is introduced; the system is composed of prescreener, discriminator and classifier. The prescreener uses double parameters-CFAR algorithm to roughly detect potential targets. Canny operator is applied in discriminator to extract edges of target, and we use Hausdorff distance transform to get the distance image, after the processing, the correlation is improved remarkably. But because edge detection and distance transform increased the computational cost, during the course of matching in the third step, we suggest two kinds of optimized search strategies to speed up recognition.Some new methods are suggested in this paper: 1)A self-adaptive region segmentation method which based on the difference of mean intensity and deviations, this method avoids threshold selection by operator;2)Grid feature vector is introduced to reduce the size of intensity feature and speed up pattern matching;3)Canny operator is applied to extract edges of target, scaling and translation transform can be overcome by using the Hausdorff transform distance feature;4)We also suggested a kind of search strategy: multi-scale searching based on wavelet decomposing.