Study On High-resolution Remote Sensing Image Nearshore Ship Detection And Location Method

With the development of aerospace technology and the progress of sensor level,remote sensing data resources are more abundant,and the resolution of remote sensing image is much higher,which brings new requirements and challenges to the task of highresolution remote sensing image interpretation.Remote sensing ship target detection and location is an important research content in remote sensing image interpretation.In the military field,ship target detection and location can provide important information support for precision guidance and long-range attack,and help to improve China’s long-range strike capability in the future;In the civil field,the detection and positioning of port ships can not only provide guarantee for maritime transportation,but also monitor illegal acts such as illegal fishing and smuggling.However,traditional ship target detection and location methods can not accurately identify the location of ship target in real geographical coordinates,and other technical means are needed to locate its geographical location.This paper aims to detect ships automatically in remote sensing images and locate the geographical coordinate level.The task is divided into three steps: 1)use the image matching algorithm to match the port test image without geographical information with the reference image with clear port location,so as to determine the port location contained in the image shooting area.2)The pixel level correspondence between the test image and its corresponding reference image is generated by the image registration algorithm.3)The pixel position of the ship target in the test image is detected by the target detection algorithm,and the geographical coordinates of the ship target are determined according to the pixel level correspondence generated in the image registration stage.Aiming at the task of ship target detection and positioning,this paper studies the image matching,image registration and multi-directional target detection algorithms based on deep learning from the perspective of improving the matching accuracy,registration speed and ship target detection accuracy of high-resolution remote sensing images.The main work is as follows:1)Aiming at the problem that the number of feature points extracted by traditional image matching algorithms in high-resolution remote sensing images is few and the feature is not conducive to matching,the feature extraction network RF net based on receptive field size is studied.When detecting feature points,RF net predicts the location of feature points with scale information combined with the receptive field range of different layer features;in the feature description,the slice position is selected based on the neighborhood mask method,which effectively improves the distinguish ability of the description vector.In this paper,RF-Net is used for image matching experiments,the experimental data include airport remote sensing images and port remote sensing images,and the experimental results show that the features extracted by RF-Net algorithm have high accuracy when used for remote sensing image matching.2)The detector-free image registration algorithm Lo FTR is studied,and a registration strategy for large and high-resolution images is proposed.High-resolution images have time phase differences,and the size differences between different images are large.Directly registering two images requires a lot of computing resources and the registration accuracy is not high.The image registration strategy based on coarse registration region constraint proposed in this paper.Firstly,registers between the down sampled remote sensing images to determine the image overlap region,then extracts evenly distributed slices in the high-resolution image overlap region for fine registration,and finally maps the registration relationship of all slices back to the original image to obtain the registration relationship of high-resolution image.The experimental results show that the proposed registration strategy can register high-resolution remote sensing images accurately and quickly.3)A multi-directional ship target detection algorithm RSI-Center Net is proposed to solve the problems of dense target arrangement,arbitrary direction and complex background in remote sensing ship detection task.Firstly,an angle branch is added to CenterNet to detect ship targets in any direction;Secondly,the semantic segmentation branch is added,and the features of the semantic segmentation branch are fused with the features of the input detection head to strengthen the features of the foreground region and weaken the features of the background region;Finally,the attention module is added to strengthen the characteristics of target salient areas and channels and improve the detection accuracy.The performance verification and speed test of this algorithm are carried out based on HRSC2016 data set.The experimental results show that compared with other advanced methods,this method has higher detection accuracy and detection speed.