Segmentation Of The Vehicle Target In SAR Imagery

Automatic recognition of military vehicles from a complex scene is an important part of automatic interpretation of SAR images. This processing is usually divided into three stages: detection, discrimination and classification. Region segmentation of vehicle ROI is a key step of target detection and discrimination. The result of target segmentation would affect the following processing performance, such as feature extraction, target classification, and target recognition, etc. So the segmentation of the vehicles in SAR imagery is a focus of attention in remote sensing information processing society. This thesis discusses some new ideas and improved approaches of threshold segmentation and MRF segmentation of the vehicles.Firstly, the existing methods of SAR image segmentation are summarized and sorted into two main categories, data-driven segmentation and model-driven segmentation. The evaluation criterions of SAR image segmentation are also given. Then, the applications of 2D histogram and Markov Random Field (MRF) in SAR image segmentation are studied systematically in this thesis. Two new approaches, an improved 2D Otsu segmentation algorithm and a fuzzy MRF segmentation algorithm with multi-model are developed.Based on the thorough analysis of the 2D histogram, this thesis gives a new method of histogram area partition for speckle noise in SAR image and brings forward a new rule of threshold selection which fits 2D histogram texture better. As a result, a better segmentation performance of the improved 2D Otsu is reached in the computational efficiency and segmentation accuracy.As for the fuzzy MRF segmentation algorithm with multi-model, the conditional probability model and the prior probability model of MRF are studied respectively. On the one hand, the optimal statistical distributions of different scenes are analyzed. On the other hand, the fuzzy method is introduced in the prior probability model. Then the multi-model and fuzzy methods are combined to form a new algorithm which models the images more accurately. Compared with the existing approaches, it can obtain better segmentation result.