Research On Technology Of Ship And Aircraft Targets Recognition From Large-Field Optical Remote Sensing Image

With the rapid development of space remote sensing technology,remote sensing image processing technology has been widely used in national defense construction and economic and livelihood construction.In particular,target detection and recognition technology for remote sensing images has become important means for military strikes and marine surveillance.The improvement of the resolution and size of remote sensing images make the data increase sharply.Extracting useful information from such huge data efficiently and accurately becomes the key to solve the problem.In order to solve the difficulty of extracting targets from complex backgrounds,enhance the reliability of remote sensing target detection technologies,and ease the pressure of data transmission and storage,this paper focuses on the detection and identification of ships and aircraft targets in large field and high resolution optical remote sensing images.Based on the panchromatic and multi-spectral optical remote sensing images,this paper analyzes the characteristics of typical targets in remote sensing images,designs a corresponding remote sensing image target detection algorithm,and elaborates on the key technical theories of target detection,such as target potential scene extraction,target location,detection of interest areas,and feature extraction.In this paper,through the research and improvement of technologies of target detection in remote sensing image,the problem of slow operation speed and low detection rate in large field images is improved.According to the prior knowledge,the ship will locate in the ocean and the aircraft will stop at the airport.In the detection and recognition of large field optical remote sensing images,the potential scenes where targets appear are first extracted to lessen the search range.This paper extracts the potential scenes where ships and aircrafts may show up-sea-land separation and airport detection.In terms of land-sea separation,for the traditional method which only uses panchromatic images can make large errors for sea-land separation,this paper,using the characteristics of different seawater and terrestrial spectral information of multi-spectral images,proposes a main component normalized water index method to separate land and sea.To eliminate the interference of non-principal components such as thin cloud shadows,principal component transformation is performed on the four-spectral multi-spectral image,the transformed non-principal component is discarded,and then the inverse of the principal component is used to reconstruct the four-spectral segment image.Using the reconstructed multispectral image for normalized water body index calculation,the results of land-sea separation are obtained.In terms of airport detection,this paper makes use of the long straight-line characteristics of airport runways for airport testing.The algorithm improves the straight line segment density model.First,the straight lines of the lowresolution image are detected and the length and gradient of each line are calculated.Each line is weighted by its length and gradient to generate a weighted feature map.The longer length of the feature map is,the more obvious the straight line is.Then,the Gaussian distribution density is calculated for each point's neighborhood to obtain a straight line density saliency map.The area with a high value of the saliency map is the airport area.Experiments show that the method can extract the target potential scene accurately.Aiming at the problems of the large calculation for searching targets in global scope and the difficulty of target location in large field low-contrast images,an adaptive target positioning algorithm based on local feature sorting is proposed.This method can locate all potential positions of suspected targets without any missed alarms.First,the Harris feature map is obtained from the image,and the feature map is considered as a pulsed image.The pulse position is the target suspicious position.Then use local adaptive threshold to determine the positioning point.Finally,merge near points using feature of the Harris point cluster.Experiments show that this method can quickly locate the potential location of target from a large field image.At the same time,it has an inhibitory effect on the low-frequency region of the cloud.Aiming at the difficulty of extracting the complex target in the background of remote sensing image,this paper proposes a method of segmenting the target based on KL divergence and GrabCut model for target extraction.The method detects the saliency of the target slice.Then binary image of saliency map is used for the minimum external moment detection.The interior and exterior regions of the external moment are used as the foreground and background areas of the Gaussian mixture model respectively.Thus this solves the problem that GrabCut needs to select the background manually.Then,the KL divergence constraint term is added to the GrabCut energy equation,which can solve the problem of the discontinuity of the extracted region due to the grayscale unevenness.At the same time,it improves the speed of iterative convergence of the energy equation and is more suitable for remote data processing and application.Aiming at the difficulty of target recognition caused by the irregular shape of the extracted aircraft target,this paper proposes a corner point convex hull segmentation algorithm and establishes two new features.Firstly,Harris corner extraction algorithm is used to extract the target corner points.Then the convex shape of the corner points is made by using the inherent shape information of the aircraft.The target is cut into several fragments according to the vertex and centroid of the convex hull,and two sets of new features are obtained by calculating the proportion of the fragments.These two sets of characteristics are used to confirm whether the target is true.Experiments show that this algorithm can effectively reduce the effect of shape irregularity and side-shake imaging distortion on the target confirmation and improve the detection rate.In terms of hardware development of the algorithm,based on the FPGA+DSP combination platform,this paper transplant software algorithms to the embedded system.The data processing module uses FPGA to perform data preprocessing,and the algorithm module uses DSP to implement.The system adopts a multi-core parallel processing method.The parallelization and optimization of the algorithm improves the processing efficiency of the system.Experiments show that the system can satisfy the requirements of real-time processing.