Research On Moving Ship Detection And State Estimation In GF-4 Remote Sensing Image

Nowadays,China’s mature land security monitoring technology has entered the homes of ordinary people.Compared with the vast land area of China,it is equally important to establish sea surface observation and monitoring in China’s 4.7 million square kilometers of sea area.With the rapid development of satellite imaging technology in recent years and the continuous improvement of high-resolution satellite resolution in geostationary orbit,it is possible to carry out long-term and wide-ranging observation and monitoring of surface ships.Today,the detection technology based on convolutional neural network has achieved good detection results.The research of moving ship detection based on high-resolution optical remote sensing image of geostationary orbit satellite based on convolutional neural network target detection technology will have great significance in the fields of civil transportation,rescue,fishing and hunting。Aiming at the design problem of the detection scheme of sea moving ships,through the analysis of the single-stage detection method and the two-stage detection method based on the convolutional neural network,the single-stage and two-stage ship detection schemes are studied respectively.The comparison test on the ship inspection data set finally determined the two-stage ship inspection scheme with better inspection results as the research scheme in this paper.Then,starting from the goal of improving the detection effect of small target ships,this paper launched an optimization study on the three shortcomings of the plan for theshortcomings of the two-stage detection scheme in ship detection applications.First,for the problem of excessive downsampling of small targets by the original feature extraction network,this paper proposes a method of using the feature extraction network HRNet that can obtain high-resolution feature maps to achieve ship feature extraction,and through the feature pyramid structure based on HRNet Designed to achieve the preservation of ship spatial information in the process of feature extraction;secondly,for the problem of single ship size in the research data,this paper analyzes the applicability of multi-scale detection to reduce network learning difficulty and improve detection From the perspective of effect,a single-scale ship detection method is proposed.Finally,for the difficulty of ship target feature extraction resulting in inaccurate classification,this paper proposes a method of target local context feature extraction.The combined use of deep context features enhances the classification features and improves the detection effect.Through these three points of optimization,a good detection effect of improving the mAP of ship detection to more than 0.9 is achieved.Finally,this paper studies the problem of ship motion state estimation.By analyzing the mechanism of multi-channel time-sharing imaging of geostationary satellites,a method of ship motion state estimation is proposed by using the imaging time difference between different channels and the small displacement of the image.Then through the analysis of the image quality,the paper proposes to use ortho correction to improve the geometric distortion caused by the image,and uses the ship super-resolution technology tosolve the problem of insufficient resolution of the ship image.Through the optimization of these two aspects,the motion state estimation of the moving ship with small error is finally realized.