Target Detection For SAR Image With Structural Information And Statistical Features

As the first step of Auto Target Recognition (ATR) system, target detection for Synthetic Aperture Radar (SAR) image has very important implications on the fellow-up processes in ATR system, such as target identification, target recognition, target classification and so on. It has very broad application prospects to carry on target detection for SAR image in the field of Intelligence SAR Image Processing and Interpretation. And in the application of ground military reconnaissance, target detection of SAR mainly detects such artificial targets as bridges, ports, vehicles, ships, buildings and so on. Primal Sketch model is a sparse representation for structural information of image, which can depicts the singular information in an image such as point target, line target, target boundary and so on, in the form of blob and sketch. The OTSU image thresholding algorithm, proposed by NOBUYUKI OTSU, is a description for the gray statistical information of image, which segments an image with the thresholds that are properly selected by maximizing the ratio of between-class variance and within-class variance. In this paper, a structural information and statistical features based target detection method for SAR image has been proposed, which integrates Primal Sketch model and OTSU image thresholding algorithm. The main contributions can be summarized as follows:Firstly, we study regularity feature based region remark extraction algorithm that is proposed by Liu Fang and Song Jianmei, by which the region marks are obtained. Corresponding to these region marks, the potential areas of artificial targets are extracted from the original SAR image, which cover almost all of the artificial targets in the original SAR image. But it only makes use of structural information of image extracted by Primal Sketch model, and since the gray characteristics of SAR target are not sufficiently made use of, the original algorithm has some shortcomings such as high false alarm ratio and not accurate enough locating of artificial target. In response to these shortcomings, we do some post-processing on the potential areas of artificial target, which integrates the original algorithm and the OTSU image thresholding algorithm. The false alarm targets are removed and the artificial targets are located more accurately in the way of gray statistics and other features. The simulation results show that the target detection algorithm discussed in this paper can effectively detect artificial targets of various types, and produces a lower false alarm and more accurately located artificial targets.Secondly, we make in-depth analysis and discussion on the regularity feature based region mark extraction algorithm. The causes that lead to the high false alarm ratio and not accurately located artificial target are discussed. We make improvements to the original algorithm in such aspects as seed segment selection, the use of non-recursive strategy, and adding growing rules, etc. Specifically, these improvements include setting up regularity and regularity ratio threshold for the seed segment, taking a non-recursive strategy which choose the original seed segment as benchmark in the process of region growing, and adding regularity requirement for the segment which will be added into the target region, etc. The effectiveness of these improvements is verified through the comparison of simulation results.