On-orbit Ship Detection In Satellite Remote Sensing Images Based On Local Features

With the essential military use and unique political significance of large ships(e.g.,aircraft carriers),promptly detecting and following these ships can help to formulate the national security strategies and develop the national economy As the requirement of real-time detection and the limitation of satellite-to-ground data transmission,it is inevitable to realize ship target detection of remote sensing images in orbit.However,on-orbit detection faces several challenges,such as the performance limitation of on-orbit satellites,the high resolution,large scale and small target characteristics of remote sensing images.According to the characteristics of high-resolution remote sensing images and on-orbit ship target detection tasks,this thesis analyzes existing target detection methods and evaluate them with experiments.It proposes an algorithm that uses accelerated SNIC superpixels to segment high-resolution remote sensing images and realize target pre-positioning on the premise of ensuring the integrity of the target.Utilizing local features and variable threshold neighbor check to realize target detection in the pre-positioning area.The computational efficiency and detection accuracy of the algorithm are compared and verified with experiments.The research works and contributions in this thesis can be summarized as follows:1.In terms of superpixel segmentation,this article compares state-of-the-art superpixel algorithms and selects the SNIC algorithm,which has the best balance with performance and efficiency.To improve the segmentation speed,the image is segmented after down-sampling.We accelerate the original pixel distance calculation method by transformation,formula deformation and approximation.The floating-point calculation data of the original algorithm are converted into reshaping calculation through normalization and translation.The updating strategy of super pixel center is improved,and the segmentation efficiency is further enhanced.The super-pixel edge is expanded to become a super-pixel structure with overlapping edges,and the super-pixel expression ability is improved.Fast filtering is carried out on each superpixel to determine the superpixel that may have a target and map it to the original high-resolution image.Reducing the target detection range from the full image to individual superpixel regions greatly reduces the detection range.2.For the selected super pixel region,the Speeded up robust features are used to extract the feature points of the region.We compared and evaluated existing feature point matching algorithms.Ratio-test algorithm has most balanced accuracy and time complexity under the application scenario in this thesis.By calculating the matching distance distribution between pairs of points,we fit the distribution with a normal distribution and generate a variable threshold based on it.Then,we proposed a variable threshold neighbor distance ratio algorithm.At the same time,the piecewise linear function is used to fit the probability density function to reduce the calculation amount.Design experiments verify the feasibility of the fit.A dual neighbor distance check algorithm is proposed to solve the problem of leak detection caused by the multi-dimensionality of feature point descriptors,the variability of target features and the insufficient use of neighbor distance information.Finally,we realized real-time detection with high accuracy.3.Combining the accelerated superpixel segmentation,mapping and variable threshold neighborhood ratio matching algorithm,we design ship detection algorithm and realize its software.The detection performance and detection time of the current mainstream target detection algorithm and the algorithm in the satellite platform simulation environment are compared.The experiments demonstrated that the algorithm in this thesis has the best performance under the premise of the performance and real-time requirements of the on-board platform.The operation efficiency of the algorithm can meet the real-time detection requirements of the on-board platform and can accurately detect large ship targets in high-resolution remote sensing images.The algorithm designed in this thesis has important research value and worthy of being further studied.