Study On Feature Extraction Methods Of SAR Images

With the rapid development of synthetic aperture radar(SAR), feature extraction methods of SAR images have become a hot research topic. There are many features in SAR images, such as edge, corner, texture, straight line and shape, in which edge and straight line are the most common. Hence, this dissertation focuses on how to efficiently, rapidly and reliably detect edges and straight lines from SAR images.For SAR edge detection, the contributions of this dissertation are as follows.1. The dissertation proposes an edge detector with low false positive detection rate. Most of commonly-used SAR edge detectors always include heavy tails in image space. And, they often have high first side lobes in frequency space. Heavy tail and high side lobe are the important reason leading to high false positive rate. Furthermore, multiplicative speckle noise has a significantly influence on the quality of SAR images, which increases the difficulty of edge detection. Hence, edge detectors based on single pixel cannot obtain satisfactory detection results. The proposed detector can fully utilize the integrative information of multiple adjacent edge-pixels, which can effectively reduce the number of false edge-pixels. More importantly, the proposed detector not only includes short tail but also has low side lobe. Experiment results demonstrate that the proposed detector can markedly decrease the false positive rate.2. This dissertation suggests an edge detector with high true positive detection rate. Edges in real scenes are complicated. Edges formed by the regions with high(low) contrast are referred to as strong(weak) edges. Comparing with strong edges, it is more difficult to correctly detect weak edges. Hence, weak edges are usually ignored by commonly-used SAR edge detectors. For effectively extracting weak edges, we propose an edge detector of SAR images using crater-shaped window with edge compensation strategy. By using crater-shaped window, the proposed detector can efficiently reduce the number of false edge-pixels caused by the noise. Furthermore, by using the edge compensation strategy, the proposed detector can detect almost 70% weak edges. Experiment results demonstrate that the detector with the edge compensation strategy has high true positive rate.3. Under the influence of the speckle, boundaries in SAR images are always fuzzy and unclear. Theoretical analysis demonstrates that traditional ratio-based edge detectors cannot obtain unbiased edge detection results, if the detected edges are fuzzy. That is to say, when a traditional ratio-based edge detector is used to detect a non-ideal edge, there is a bias between the location of the detected edge and its real location. Difference-based edge detectors can correctly detect the non-ideal edges, but because of the speckle, they do not keep the constant false alarm rate(CFAR). In this dissertation, an unbiased edge detector with CFAR is proposed for detecting non-ideal step edges in SAR images. The proposed detector includes a difference operation and a ratio operation. Theoretical analysis results show that the difference operation provides unbiased localization for non-ideal edges, and the ratio operation keeps CFAR property under the influence of the speckle noise. Both objective and subject experimental results show that the proposed detector can provide unbiased edge position and obtain closed skeleton edges.For SAR straight line detection, the contributions of this dissertation are as follows.1. Traditional straight line detection methods usually detect straight lines from binary edge maps. That is to say, the detection result of a traditional straight line detection relates on the quality of a binary edge map. Due to the influence of the speckle, there are more or less false edges, distorted edges and even biased edges in binary edge maps. These false, distorted and biased edges largely affect the detection performances of traditional straight line detection methods. Hence, in this dissertation, we present a line extraction method. It directly detects the straight lines from the image edge fields of SAR images, rather than from binary edge maps. For both synthetic and real world images, our method achieves a good performance in true positive detection rate and detection accuracy.2. In general, the computational complexities of commonly-used straight line detection methods are connected with the scene complexity of an image. Based on the funnel transform, this dissertation introduces a new straight line detection which can efficiently and rapidly detect straight lines. The computational complexity of the proposed method is only connected with the size of the detected image. On the basis of the slope-intercept line equation, our method makes the complex straight line detection problem be changed into an easy one for identification of local maximum points(peaks). The parameterization information of a line can be uniquely recovered from the coordinates of the corresponding peak. The proposed method directly performs on grayscale images. For straight lines formed by ridge-typical and step-typical edges, our method can uniformly maps them into some sharp peaks. Theoretical analysis and experimental results demonstrate that the proposed method has advantages including lower computational complexity, higher detection precision, stronger anti-occlusion and noise robustness.