Research On Ship Detection Algorithms In Harbors Based On Optical Remote Sensing Images

Optical remote sensing images have played an important role in civil and military use owing to its outstanding advantages of visual interpretation,intuitive interpretation,and easy interpretation.In recent years,target detection and recognition in optical remote sensing images have been researched by many domestic and foreign research institutions and scholars,and have made a series of progress.This paper focuses on the accurate detection of ships within the optical remote sensing image port.Compared to the sea surface ship detection,ship detection more difficult in the harbor.Because there are many problems such as ship adhesions,ships and islands adhesions,and the confusion between ships and the ground construction of ports.These questions are still open and challenging issues without effective solutions.On the one hand,this paper combines the port prior knowledge,constructs the port template to complete the accurate detection and extraction of the port,and simplifies the complicated problem of ship detection in the ports as a surface ship detection problem,which avoids the influence of port architecture on the detection algorithm.On the other hand,on the basis of this,combined with the powerful learning ability of deep learning,an in-ship detection and detection algorithm based on deep neural network was proposed and satisfactory results were obtained.The main work and contributions of this article are as follows:(1)An optical remote sensing image registration algorithm based on hybrid features is proposed.For high-precision high-resolution optical remote sensing image registration,due to changes in shape,color,and texture,feature points have large randomness,low matching rate,high false matching rate,and are susceptible to environmental noise.We propose a high-precision optical remote sensing image registration algorithm.The algorithm extracts the ROEWA-Harris feature from the remote sensing image,and then combines the Lo G-Polar and SIFT mixed description operators to achieve accurate registration of the port image using the mixed features.(2)For optical remote sensing port images,based on the characteristics of the port remote sensing image,combined with the algorithm proposed in(1),a high-precision port segmentation and extraction algorithm is proposed.The algorithm firstly creates F(Feature)templates and B(Binary)templates based on port prior information,and the function of the F template is to achieve accurate registration of the detection image to the port template.The role of B templates is to achieve accurate segmentation and extraction of harbor sea areas.In the actual detection process,firstly,the port area of interest is roughly located by the latitude and longitude coordinates contained in the remote sensing image data,and then the precise registration of the detected image to the B template is performed according to the F template corresponding to the port,and finally,the high precision segmentation of the port is completed according to the B template.The analysis of the experimental results shows that the size and direction of the images are the same as those of the template image,with accurate sea-land segmentation,which is more conducive to the high-precision detection and identification of ships in subsequent ports.(3)An algorithm for the detection of ships in the harbor based on convolutional neural network is proposed.For the problem of optical remote sensing ship detection,convolutional neural network method is used to carry out ship detection on the optical remote sensing port image,input remote sensing image,and output ship target number,area,and classification confidence.This paper studies the construction and calibration of training data,neural network structure and parameter optimization,CNN-based ship detection algorithms and performance evaluation of algorithms,and analyzes the performance of the algorithm through specific experiments.