Research On Typical Linear Target Extraction From Remote Sensing Image Based On Feature Level Fusion

Developing the automatic intelligent target extraction technique is the effective way to release the pressure of data applications. In order to settle the mutual questions on the target extraction and deal with the variety of targets and sensors, typical linear targets (such as airports, roads etc) are selected as the main research object. By integrating the geometrical and statistical features in multisensor image, the fused target will be accurate and reliable.In the field of image understanding, region and edge are the basic features describing the objects. The technique of feature extraction includes region segmentation and edge location. So, the work in this thesis focuses on three aspects: (1) Analyzing the geometrical structure feature and spectral feature of typical linear targets in the multispectral image and the panchromatic image. By introducing the knowledge expression and illation, developing the idea target model and detecting the target from one optical image. Then, Based on the the redundancy and complement of the multispectral images, the research on merge strategy of fused target extraction is hold on. (2) With the help of texture analysis in the SAR image, choosing the best texture operator and segmentation appraise rule, by which the whole target topology and region is easy to acquire. Seeking the better fusion strategy to solve the problem of merge the target region segmented from the different polarization mode of SAR sensors. (3) Proposing the model of step edge orientation in remote sensing image and finding the edge associate method between the different special resolution edge images.The conclusions and originalities of this dissertation is listed as following: (1) The multilayer target extraction model application on the typical linear target from the multispectral images is presented, which include three stages according to the local-to-whole: ideal road snippet extraction, region growth and target feature fusion. Where, in the first stage, the parallel beeline is replaced by the parallel curve defined with the equality of the distance on direction of the normal. It is closer to the real road than the beeline model. Moreover, the D_S evidence theory is introduced in the fusion stage. It imposes the redundancy and complement of the multispectral images to improve the detection probability, especially to the target that is made up of different materials.(2) Based on the integration regions from SAR image with edges from optical image, a new target extraction algorithm is proposed. During the course of region segmentation, it is proved that the texture inertial moment is the best operator, and the fuzzy C-means clustering method based on mahalanobis distance of feature sets is highly effective.(3) To solve the fusion question of linear target region detected by different imaging mode SAR sensors, the equivalent noise looks is expressed as the weight factor of the fuzzy grade of membership. The results illustrate that this algorithm can depress the false alarm rates and hold the integrality of topology structure.(4) The subpixel edge orientation technique based on the improved gray moment is discussed. Even if there are different original phases in the image captured by the same sensor, it can ensure that the positive and negative step edges shift on the same direction, namely the width of the object is constant. This method supplies the steady edge data to the fused edge.(5) Feature vector, which includes sign of edge, central angle and the distance from chord to the circle center, is served as the role of feature association. With it, the model of edge fusion is approved. The example of road extraction in different special resolution optical images shows the efficiency of this fusion method.