Target Detection In Optical Remote Sensing Images With Complecated Background

Target detection is to detect and trace the interesting targets in the satellite images by image processing techniques. It is one of the important applications in remote sensing technique. With the various information obtained from the satellite at any time and under any weather, the target detection is widely used in national planning, disaster monitoring, and military reconnaissance etc. Recently, with the development of remote sensing, the target detection technique has greatly improved especially in the aspect of image resolution. The resolution of the optical remote sensing images is as high as 0.6m per pixel. In such high-resolution optical remote sensing images, the contours and the textures of the targets are clearly observed and informative for target detection. In a word, the study on the target detection in optical remote sensing images has great significance in aspects of theorical research and practical application.In order to solve the robust problem of the remote sensing detection in complicated background, this paper proposed solutions on the color-texture interference problem, the rotation-scale variance problem and the shape-similar distractor interference problem. The applicability and the validity of the proposed method is proved and discussed through theoretical and practical experiments. The main innovations of the proposed method are as follows:(1) To reduce the interference of the color-texture interference in complicated background, a geometric part-based model is proposed. Making use of the geometric characteristic of the profile of the rigid objects, a geometric atom dictionary is built. Under the framework of sparse representation, the contour of the rigid targets is represented by a small number of geometric atoms, which are taken as the parts of the geometric part-based model. To describe the space relationship of the geometric part atoms, an ordered-link structure is designed based on the pre-defined weights of the parts. Based on the ordered-link structure, a hieratic detection method is proposed. The experimental results demonstrate the adaptivity of part selection, the robustness to color, texture and the complicated background interference, and the efficiency of the part detection. The propose method has a good performance in detection rigid targets in complicate background.(2) Objects are usually in different directions in remote sensing images. To improve the tolerance of the orientational variance in the complicated background, a rotation invariant target detection method based on the radial-gradient angle (RGA) and the consistency voting is proposed. First, the radial-gradient angle, which is a rotation invariant feature, is defined to describe the edge pixels. Second, pairs of related pixels are searched by the pixel matching based on the RGA. Third, a set of possible principal directions of the target in each detection window are estimated by pairs of related pixels. Finally, all possible principal directions cast vote to obtain the detection score and the principal direction of the detected target. The experimental results indicate that the proposed method can precisely detect the target and estimate the principal direction of targets under the background with shadows and occlusion. It outperforms other existent detection methods in complicated background.(3) In order to solve the scaling variance problem and eliminate the interference of the shape-similar distractor in the complicated background, a detection method based on the pose weighted voting is proposed and applied on the inshore ship detection issue. To achieve rotation-scale invariance, the target pose is defined and used in both of the estimation stage and the voting stage. To increase the discrimination between the ships and the distractors with rectangle shape, each edge pixel on the ship template is weighted by a weight matrix, and the weight for the V shape structure of the ship head is set larger. Based on the weight matrix, a pose weighted voting process is designed to obtain the pose and the votes of the detected ship. Moreover, to distinguish the ships from the distractors which have the V shape structure, an outline continuity factor is defined to evaluate the continuity of the detected target line. The detection score is refined by the outline continuity factor, leading to more precise results. The experiments discuss the selection of the weight and the outline continuity factor, and demonstrate the effectivity and the robustness of the proposed method by comparing the inshore ship detection results with other state-of-the-art methods.