Researches On Konwledge-Aided SAR Target Indexing And Feature Extraction Technology

Automatic Target Recognition(ATR) using SAR images is an important topic in SAR applications. The main challenge of SAR ATR roots in variability of targets and complexity of environment, under which limited information embedded in SAR observation is far from sufficient for target recognition. Other sources of extrinsic knowledge and aided information should be introduced to cope with the challenge. Focusing on this point, some related issues are discussed.After reviewing the status and problems of SAR ATR, the thesis points out that Knowledge-Aided SAR ATR is a feasible way attacking the dilemma of SAR ATR, and a hierarchical, adaptive information processing flow is further proposed by introducing knowledge about targets, environment, sensors and domain expert. “Target indexing” is a key element in the Knowledge-Aided SAR ATR, and its task is to locate hot spot regions in the high dimensional hypothesis space, and provide a coarse hypothesis list for subsequent finer and complicated processing. By paying attentions on Knowledge-Aided and Extended Operation Conditions(EOCs) processing, the subject of Knowledge-Aided SAR target indexing and feature extraction is focused.Knowledge-Aided SAR chip image segmentation is firstly studied in the chapter two. For SAR images of ground vehicles, shadow is also the interested region. Existing segmentation methods are mainly based on different grayscales between target and background, and local connectivity between adjacent pixels, and therefore loses many meaningful target and shadow regions. By introducing the prior knowledge of spatial connectivity between target and shadow, which is modeled as Spatial Relation Potential Function(SRPF), a new method called SRPF-MRF is proposed to retrieve more complete target and shadow. Compared to MRF segmentation, SRPF-MRF imposes only an extra prior probability for each pixel, and its incremental computation is few, thus statisifies target indexing well in computational efficiency.Scattering centers extraction in the context of target indexing is investigated and researched in the chapter three. Firtly attributed scattering centers extraction and high resolution complex-imagery scattering centers extraction are investigated and pointed out to be not suitable for target indexing because their disadvantages in computational efficiency, robustness and automation. Peak extraction statisfies these demand very well, however, it fails to consider distributed scattering centers, and loses many meaningful scattering centers. By illustrating attributed scattering centers and its approximate equivalence with ideal point scattering centers in real SAR image, a new scattering center model is formulated. According to this new model, a CLEAN method for scattering centers extraction is proposed. The proposed method can evidently mitigate the missing detection of scattering centers. Futhermore, with the schema of “CLEAN”, the proposed method is computational efficient, robust and automatic, and statisifies target indexing well.In the chapter four, aiming at gound vehicles, and according to their simpilified schematic model: cubiod, we derive that ground vehicles have rectangle outline in SAR images. The rectangle outline is a kind of knowledge about target model, and provides a suitable context for representing subparts of vehicles, such as tank turret, is therefore profitable for target indexing under EOCs. The critical problems that handicap the accurate rectangle outline extraction are fake objects, missing, and possible articulations during SAR chip image segmentation. By introducing prior information about shadow boundary and vehicles’ length-width ratio, a heuristic method is proposed to attack these problems, and extract accurate rectangle outline of vehicles. Experiment results on MSTAR dataset validate the the proposed method.The chapter five focuses on characteristic-substructure based coarse target clssification(namely indexing). The characteristic-substructures, and their reprenstation, extraction, predicting and applications in SAR target indexing is firstly discussed. From the pointview of Knowledge-Aided, the characteristic-substructures synthesis knowledge from target models, observations, sensors and experts, etc. With the the context of rectangle outline, and according to the characteristics of tank turret in SAR images, we detect and extract tank turrets from segmentation labels and scattering centers. The detected tank turrets are subsequently used in characteristic-substructure based indexing. Persisting scattering centers are extracted as characteristic substructure, and are used in chacteristic-substructure matching based indexing by matching the observed SAR image scatrering centers with persisting scattering centers in the context of rectangle outline. Experiments on MSTAR dataset show the validity of these characteristic substructures for target indexing.