Research On Optical Remote Sensing Detection And Recognition Algorithm Of Ship Target Based On Deep Learning

As the carrier of maritime trade and transportation,ships are also important military objects.The automatic detection and identification of ships has far-reaching practical significance.With the development of remote sensing imaging technology,ship target detection method based on optical remote sensing image is widely used.However,the traditional optical remote sensing processing methods can not accurately describe the characteristics of the ship.Due to the influence of clouds,waves,clutter and other factors,the detection robustness is poor and the false alarm rate is high.Deep learning method can obtain high-dimensional features of ships from complex background,so as to extract more detailed information.This method gets higher accuracy,which can meet the real-time needs.In this paper,a rich set of ship samples is constructed,ship detection and recognition research is carried out based on the mainstream deep learning algorithm,and the YOLOv4 algorithm is improved.The main work of this paper is as follows:(1)In order to solve the problem of lack of ship sample set,this paper constructs a ship detection data set of 3893 images,including 8098 ships of various sizes in different scenes.The image resolution is between 0.3m and 1m.Furthermore,the data set of ship type identification is constructed,which includes military ships,bulk carriers,container ships,tanker,passenger ships and so on.(2)The research of ship detection and recognition based on deep learning algorithm is carried out.The ship detection performance of YOLOv4 algorithm is the best,the AP value reaches 92.30%.The Faster RCNN is close to the AP value of YOLOv3.The Faster RCNN method can detect more ships,but the false detection bias is more,and the detection robustness deviation for small ship.The AP value of SSD is low,and the overall detection effect deviation for small ships in complex background.The ship recognition accuracy of YOLOv4 is the highest,and the m AP value is 96.76%.When detect bulk carriers,container ships,passenger ships and tanker,the accuracy of YOLOv4 is equivalent to that of Faster RCNN,SSD and YOLOv3,but the AP value of military ship recognition is 93.16%,while that of Faster RCNN,SSD and YOLOv3 is about 85%.(3)In order to solve the problem of small and side-by-side ships detection and recognition in complex background,the YOLOv4 algorithm based on attention mechanism Fca Net is constructed.The Fca Net model is embedded into the network,and 16 frequency components of low frequency are selected,and the detection and recognition results of different combined positions on the constructed ship data set are tested.The experimental results show that the three branches embedded in the backbone network output have the best effect,and the detection AP value reaches 93.04%,which is improved by 0.74%.More ships can be detected,and the detection ability of small ships and side-by-side ships is improved.In the test of multi class ship recognition data set,the map is 97.37%,and the AP value of the military ship which is difficult to recognize is 95.59%,which is 2.43% higher than the original algorithm.