Researches On Embedded Quasi Real-time Ship Detection Method Based On Remote Sensing Image

In recent years,our country has been continuously launching earth optical imaging satellites with higher resolution sensors,in which more and more image data are generated.It provides more information and brings more challenges.On the other hand,ships are important targets of monitoring and attack at sea,and their location information is an important component of national defense security.Accordingly,ship detection method based on remote sensing image has great research significanceThe essence of target detection is to extract target information by using image features.Traditional methods will greatly reduce the accuracy of ship detection when they encounter complex cloud waves and terrestrial background interference.With the continuous of deep learning,CNN has shown strong feature extraction ability and superiority in object detection.This paper uses the method of deep learning to detect the ship targets in quasi real-time based on remote sensing image information.Firstly,this paper proposed a yolov3-based fast target detection network model,which cropped the numbers and layers of network and used depthwise convolution instead of ordinary convolution to reduce network parameters and achieve network acceleration.Then,in order to evaluate the scale adaptability of the model,this paper defined and measured the OSIT.After that image pyramid was used to expand OSIT scale and realize multi-scale target detection.Next,in order to solve the problem of OSIT target training,this paper proposed to use the ROO train method.This method can also solve the problem of positive and negative sample imbalance and speed up the convergence of network training.Finally,in order to make up for the lack of generalization ability of manual cropping of ROO,this paper used OHEM method to improve training effect.Based on the TX2 embedded platform,this paper uses the self-made ship dataset and DOTA dataset to evaluate the performance of the model.The experimental results show that the speed of this method is greatly improved compared with the state-of-theart network in the case of little difference in AP.Compared with the original YOLOv3,the speed is increased by 77% and AP has been reduced by only 5.7.Compared with YOLOv2.there are also 48% increase in speed and 0.4 improvement in AP.At the same time,based on the method of ROO+OHEM,the training effect and convergence speed has also been greatly improved.