Research On On-orbit Ship Detection Technology Of Optical Satellite Based On Artificial Intelligence

In recent years,the ground observation technology in the field of remote sensing has developed rapidly,and a large number of high-resolution optical remote sensing satellites have emerged,which has promoted the application of space remote sensing technology in military,national economy,world environmental protection and many other fields.As important transportation tools and weapons and equipment,warships are particularly important to intelligently detect and identify them.The rapid determination of the number of ships and their positions in remote sensing images within a short time is of great significance for traffic surveillance and trade development.However,with the increase of the space resolution,time resolution,spectral resolution,and the width of ground observations of the spacecraft,the amount of remote sensing data has increased exponentially,but the corresponding remote sensing data processing technology has progressed slowly.The detection and recognition of targets is still stuck in simple data processing or even human eye observation.The processing efficiency is very slow,which is also the main contradiction restricting the application of space remote sensing technology.Therefore,research on fast and effective remote sensing image ship detection technology is of great value to aerospace industry.Based on the above analysis,this paper proposes an on-orbit fast ship detection mode for optical remote sensing satellites,which uses deep learning-based artificial intelligence technology to quickly process remote sensing images in orbit.The ship target information can be obtained through satellite networking or relay satellites,and sent to the ground quickly.In this context,this paper has carried out research on the detection technology of on-orbit ships based on artificial intelligence for optical remote sensing images.The research is summarized as follows:1.The salient target is the area in the image that can most interest people and can best represent the content of the image.Visual saliency has been widely used in computer vision tasks such as target detection and tracking,but for remote sensing images with extremely high background complexity,traditional saliency algorithms are difficult to extract valid targets.In this paper,the research significance of visual saliency,method principle,basic model architecture and its application in intelligent detection of optical remote sensing image targets are studied in depth.This paper proposes a neural network model(A-FPN)based on saliency features which uses a saliency mechanism to guide the learning path of a convolutional neural network model,gain effective region features,suppress background information,and improve network performance.In addition,taking into the requirements for the calculation amount of the algorithm in orbit applications,this paper carried out lightweight cutting and optimization of the model for the A-FPN network,which improved the practical application value of the algorithm.Experimental results prove that the model can accurately and quickly extract effective target information from complex backgrounds,and is very suitable for ship target detection in remote sensing images.2.Optical remote sensing satellite on-orbit ship detection technology needs to consider the computing power and power consumption of the satellite platform,so the satellite on-orbit ship detection algorithm must have stability,accuracy and simplicity at the same time.The size of remote sensing images is very large.It is difficult for common target detection neural network models to detect targets directly.Redundant cropping of remote sensing images is required,and then processing is divided into blocks.This will cause the time complexity of the algorithm to be severely restricted by the image size.Cropping or design techniques to reduce the complexity of the model can not fundamentally solve the problems caused by image scale.Therefore,this paper proposes a ship detection method based on artificial intelligence.First,the integrated learning method is used to realize fast search of ships without scale restrictions,which solves the problem of limited size of remote sensing images,and then designs a deep convolution neural network for ship identification.Finally,in order to reduce the calculation time complexity of the algorithm,a model optimization method is used.Based on MobileNet,the model is cut and optimized to ensure the accuracy of ship detection.In this case,the complexity of the algorithm is reduced and the usability of the model is improved.3.In the ship target search phase,this paper proposes an improved AdaBoost algorithm.In order to extract all ship target area slices,this article adjusts the sample weight update rules of the AdaBoost algorithm.The new AdaBoost model is not only focused on the wrong classification samples,but based on the positive samples That is,the classification of ship targets adjusts the training direction of the model in real time to reduce the rate of missed detection during the ship search phase.In addition,for ship detection in the port area,considering that artificial buildings have a large interference with machine learning algorithms,this paper proposes a fast sea and land segmentation method based on Harris corners,which does not require manual marking and auxiliary information,and has high robustness.4.Taking into account that after the ship search,there will still be interferences similar to the ship,the ship identification algorithm needs to be designed to reduce the false alarm rate of ship detection.Aiming at the imaging characteristics of ship targets in remote sensing images,this paper proposes a local feature extraction method based on convolutional neural network,and then uses multi-feature fusion technology to design a neural network model with high recognition accuracy to achieve coarse to fine rapid ship detection and recognition of optical remote sensing images.5.Optical remote sensing image fast on-orbit ship detection has important practical application value.After the algorithm is designed,its feasibility needs to be verified by actual engineering.Based on the on-orbit processing algorithm of the optical remote sensing satellite and the design of the entire star platform,a spaceborne ship detection and processing architecture based on high-performance commercial FPGA and GPU is proposed.The FPGA is responsible for receiving the front-end data of the camera and transmitting it to the back-end processing.The core module of the back-end processor is a multi-core GPU,which accelerates the convolution operation in the ship detection algorithm.Finally,the intelligent ship detection system of this subject was installed on GF-03 A satellite for on-orbit verification,and the experimental results show that both the algorithm and the designed hardware platform can meet the requirements for fast ship detection of optical remote sensing satellites in orbit.To sum up: This article conducts in-depth research on the theoretical knowledge involved in the automatic detection and identification of on-orbit ships in optical remote sensing images,analyzes the main problems and challenges facing on-board orbit ships detection,and proposes corresponding artificial intelligence-based technologies The research results of this paper have greatly improved the capabilities of optical remote sensing satellites in the national economy,disaster early warning,and our defence construction,and have important reference significance in the aerospace field.