Research On Automatic Detection Of Objects With Stable Structure In Large Remote Sensing Images

Automatic targets recognition in high resolution remote sensing images is very meaningful,especially in the military filed,such as terminal guidance of aircraft. It can provides more accurate information of targets. However, Due to the variable sizes, colors,orientations and complex background, object detection is still a difficult task in remote sensing images.In this paper, we try to provide a training-free,generic,efficient method for object detection which has stable structure in high resolution remote sensing images.We divide it into three steps : pre-processing,feature extraction,detection,and test it on two types of typical targets,aircraft and oil depots.Finally our method achieves an outstanding results.In the step of pre-processing, our goal is to remove out some homogeneous backgrounds. In this step, three conditions should be met : rapidly, priori-information independent,color information independent. So we chose saliency detection method based on phase spectrum. Also, in order to improve the result of traditional method of saliency detection based on phase spectrum,we introduce the spatial competition to every points,lower the impact of texture of backgrounds. Finally combine with the MSER segmentation methods,we almost get all regions of interested which may include the targets.In the step of feature extraction, We analysis the advantages and disadvantages of object detection methods based on some commonly used features in detail,and provide a new feature base on local regress kernel.We compare our feature with some features which are widely used, including HoG features, LBP features, Haar Like features, prove that our features can get the best description and distinguishing ability.to our targetsIn the step of object detection, We choose the matrix cosine similarity to measure the difference between query image and target image according to the characteristic of our feature descriptor. Moreover,we build an simple, locally data-adaptive decision networks.Finally we give the methods to handling variation in scale and rotation.In the end of paper, we evaluate the whole algorithm in detail and analysis the adaptability of multi-targets, multi-scales and multi-viewpoints.