Small Target Detection Technique For Optical Large Scene Image

With the development of aerospace,the ability of optical remote sensing satellite to observe the earth is improving,and the spatial resolution of remote sensing image is also growing.Now,how to extract important information from remote sensing image and apply it to military defense and economic construction has become a research hotspot of remote sensing image.As an important target of marine monitoring,the rapid detection of ships in a large number of image data can play a huge role in the military and civil fields.However,the optical remote sensing image will be affected by the interference such as weather,cloud and fog,sea waves and so on,which makes the performance of the detection method decline,and the ship target occupies a small picture in the remote sensing image.How to make the algorithm focus on the target area and locate and detect the ship target on the sea quickly,reliably and accurately from the complex scene is still a difficult problem.Combined with the imaging characteristics of remote sensing images,the methods and problems of ship detection in optical remote sensing images are deeply studied.In this paper,through the research on the key technologies of ship detection,such as image defogging,sea land separation,candidate target region location,target feature extraction and classification,the small target(ship)location and detection in optical large scene remote sensing image is completed under the background interference of sea waves and clouds.The main contents of this paper are summarized as follows:This paper analyzes the characteristics of optical remote sensing image and ship target,and studies the key technologies,which provides a theoretical basis for the ship detection algorithm of remote sensing image proposed in this paper.Making full use of the multispectral information of remote sensing image to complete fog removal and land sea separation.Aiming at the problem that remote sensing image is vulnerable to the interference of cloud and fog,which leads to the lack of information,based on the atmospheric scattering model,this paper analyzes the spectral characteristics of cloud and fog,and puts forward a method to estimate the transmittance distribution map based on cirrus band.The result is more accurate,achieves better defogging effect,and shortens the operation time.In view of the problem of large error and easy to be interfered in the traditional gray-scale statistics for sea land separation,by studying the different spectral characteristics of sea and land,using the improved water index and morphological filtering method,a good sea land separation effect is achieved,at the same time,it can remove the thick cloud interference and reduce the interference of the follow-up algorithm.In order to quickly locate and screen suspected ship targets in large sea area,a saliency algorithm based on spectral residual algorithm,combined with ITTI model and gray corrosion operation is proposed.The fast positioning of ship target under wave interference is realized,and the complete and clear contour of the target can be obtained.Furthermore,aiming at the problem that the ship target with wake will appear in the extracted target slice,combined with the distribution characteristics of the ship and its wake,the grabcut method is used to segment the slice again to obtain the accurate target.The shape,texture and gradient features are used to describe the ship.And then a ship target classification algorithm based on Ada Boost classification is proposed,which can effectively eliminate the interference in the target slice and obtain high detection rate.This paper also uses DSP + FPGA hardware platform to optimize the transplantation of detection algorithm,and give full play to the characteristics of DSP + FPGA to improve the processing speed of the algorithm,which has the application potential of on orbit target detection.