Research On Key Technology Of On-Orbit Ship Detection In Remote Sensing Satellite

With the continuous development of remote sensing earth observation technology,the previous satellite working mode of "data acquisition,data transmission and data processing" has gradually revealed the disadvantages of poor timeliness and high manual dependence.The detection and identification of targets are still stuck in simple data processing and human eye observation,which is inefficient.This is also the main restriction on the application of space remote sensing technology today.This thesis proposes the in-orbit ship detection mode with domestic AI chip as satellite platform and deep learning detection technology.When obatining the remote sensing images and obtaining the target information,the results are quickly sent to the ground terminal through the satellite networking or the relay satellites.This mode can greatly improve the real-time performance of satellite data and promote the practical application of remote sensing technology.This thesis studies the detection task of ships in remote sensing images.The deep learning detection model based on convolutional neural network is designed by the characteristics of remote sensing image ship targets.The thesis mainly includes the following three parts:I.A ship dataset of 588 remote sensing images is made in this thesis.By analyzing and summarizing the ship objectives,three characteristics are large width,small target and sparse distribution.According to the characteristics of large width,this thesis proposes that the slice size should be appropriately increased to reduce the overlap rate and designs the corresponding section size calculation formula through experiments.According to the characteristics of small targets,this thesis proposes that large size features should be selected as shallow size features for prediction,which is more suitable for detecting small size targets.According to the sparsity characteristics,analyzing the detection process of one-stage and two-stage detection algorithms shows that two-stage detection is more suitable for sparse target detection tasks in remote sensing images.II.This thesis proposes an enhanced two-stage object detection algorithm.Compared with the ordinary two-stage detection algorithm,the model has the characteristic of "Onepoint weakening,Three-points enhancing"."One-point weakening" refers to reducing the network depth and channel dimension in the first stage and weakening the feature extraction capability."Three-points enhancing" are using of shallow features to detect small targets,using the YOLOv5 algorithm to extract network candidate box generation quality,and optimizing the R-CNN structure to enhance the feature extraction ability of the second stage.III.With Yulong 810 A chip as the simulation test platform,the trained model weight file is transformed to generate the corresponding C++ engineering code.Add slice operation in preprocessing and add coordinate reduction in post-processing.Finally,the algorithm design with remote sensing image as input and remote sensing pixel coordinate as output is realized.The final improved two-stage detection model achieved a detection accuracy of 98.3%on the ship dataset.Processing 2528 × 2528 size pictures only cost 24 ms,which is a 5percentage points increase in accuracy and a 22% increase in speed compared to YOLOv5 s.The experimental results verify the effectiveness of the model algorithm in the remote sensing ship target detection scenario and performs well in precision and speed.