Vehicle Target Detection In Very High Resolution SAR Imagery

Synthetic Aperture Radar(SAR) is an all-time image collector with high resolution. It is widely used in many kinds of fields such as national defense, natural disaster monitoring, ecological environment monitoring and so on. The application of SAR image includes segmentation, classification, target detection, target recognition and so on, among which, target detection is of great importance. Especially in military fields, target detection in SAR image supports many aspects such as destroying targets, precision guidance and so on. Recent years, researchers have made great advances in the field of SAR. But with the enhancing of the resolution of SAR image, traditional methods can not satisfy the needs of high resolution SAR image processing. Specifically, the three aspects followed should be emphasized: first, different from targets in low resolution SAR images, which are consisted of only one or several pixels, targets in images collected by high resolution SAR is of great difference, most of which are of an area and are with distribution characteristics. So we cannot use traditional CFAR algorithm to detect the targets in high resolution SAR images. Second, because of the characteristics of targets in high resolution SAR images with certain heights have a shadow area at a certain direction, which may leads the imperfection of the targets. Third, as the resolution of SAR image becomes increasingly high, characteristics of SAR images become increasingly more, which make curse of dimensionality of characteristics and following more time consuming in target detection.In this article, to solve these problems, we propose some solutions.First, we proposed a mothed for target detection in high resolution SAR image with combining SAR classification and fitting characteristics. In this article, as the PDF of targets can be fitting and every target have some about five thousand Pixels or more, which makes the targets area can be regarded as a class, such as Grass, runway and construction so we can classify each area in high resolution SAR image, and combining the characteristics of target we can detect the targets.Second, we will compensate the targets with imperfection. Airborne SAR image have an area with low intensity with each area with height which make the detection result imperfection. In this article, I represented a method to compensate targets with make the low brightness as a characteristic and modeling the joint area. Experiments show this method has achieved good results.Third, we proposed a method named sparse characteristics hierarchical representation to solve the complication in characteristics representation. In the first problem, to respect the features of SAR image is also a difficult problem. Because of the high resolution, features in SAR images is tiny. Traditionally, rich features favor the target detection, but so many features in the high resolution SAR images may lead high complexity in the optimization procedure. To solve this problem, we propose a method using over complete dictionaries. We use the dictionary to sparse respect the features. In the real process we find that the traditional distribution function cannot fit the SAR image well: Four order distribution can fit well with the shadow but it cannot fit other parts of the image. Gaussian distribution can fit some background regions but is useless with the target and the shadow. So we propose a mix distribution to solve the problem and get good results.