Research On Ship Detection And Identification Based On Deep Learning

With the strengthening of my country’s national strength and the irreversible trend of globalization,the volume of foreign trade is in a state of continuous growth,of which maritime transport accounts for a large proportion,and ensuring the safety of maritime transport has become the main responsibility of the current maritime monitoring department.In addition,China,as a major maritime country,has a long coastline,extremely rich mineral and oil resources in the sea,and is adjacent to many countries,so the maritime homeland security is threatened by many threats..For this reason,it is necessary to improve marine monitoring technology.At present,the detection and recognition of marine targets based on deep learning is a major research hotspot in the field of marine traffic safety.Aiming at the problem of how to detect and identify marine ship targets quickly and accurately,using the advantages of deep learning technology can provide a basic theory for improving the accuracy of marine ship detection in the current era of big data.This paper adopts YOLOV3(You only look once),a target detection algorithm based on unified regression classification,and improves and optimizes the algorithm to make it more effective to complete the detection of ships at sea.The main research work of this article can be divided into the following points:Firstly,select two types of ships and make corresponding data sets of ships at sea.Due to fewer images are acquired,image processing techniques,such as flipping,zooming,contrast and brightness adjustment,are used to increase the number of samples,with particular emphasis on increasing the ship’s images taken from the top view to ensure that the model can more accurately predict the target at this angle.After data sample enhancement,3400 samples were obtained.Then,the K-means is used to cluster the ship’s width and height in the data set,and the common ship’s size is selected as the prior box knowledge,which is helpful to the faster learning parameters of the model.In this experiment,the cluster number K belongs to the integer of [4,24],and the average IOU and size of boxes under different K values are recorded,so as to select the appropriate K ship width height data according to the average IOU value in the later stage,and apply it to the detection process of each scale of the detection network.Secondly,on the basis of YOLOV3,the FPN-like idea is used to increase the detection scale,and the third scale of the original model is fused with the feature map of the same size of the lower layer,and then sent to the detection network.The large-scale feature map has small perception,and its features tend to be local and detailed information,which helps to improve the detection accuracy of small targets.At the same time,the upsampling stage is optimized.Although the nearest neighbor interpolation method adopted by YOLOV3 is simple,the restored image edge has jagged and mosaic,and the feature information is incomplete.Therefore,bilinear interpolation algorithm will be used to restore the image feature information as much as possible without affecting the reasoning speed of the model in this experiment.Experiments show that bilinear interpolation performs well in preserving image edge information.Thirdly,In the multi-scale fusion stage of the network,the fusion method of add number of channels is abandoned,and the DCA fusion strategy is introduced,which makes the difference between the input high-level and low-level feature maps more prominent and maximizes the correlation.Then concat operation is performed on the two features transformed by DCA.The experimental results show that the fusion efficiency is improved,the interaction between high and low level feature information is more sufficient,and the multi-scale detection is added to improve the detection accuracy of ship.Fourthly,the improved YOLOV3 model is applied to ship detection.Although the detection accuracy of the model is improved,the predicted boundary box is far away from the real box,and the prediction of position information is not accurate.In order to solve this problem,based on the improved YOLOV3 model,GIOU index is introduced to calculate the loss of box.The experimental results show that the model with GIOU index has a certain effect on the prediction of box’s positional correction.The loss function curve before and after the improvement shows that although both of them are in convergence state,the improved loss value is low as a whole,and the loss oscillation amplitude between the same iterations is smaller than the original loss,which indirectly indicates that the robustness of the model is enhanced.The detection accuracy of the improved model based on GIOU loss is higher than before.Finally,the improved YOLOV3 based on GIOU loss is compared with the YOLOV3 model.Under the two models,category P-R curve analysis,analysis of model reasoning speed and detection accuracy are carried out.The experiment shows that although the speed has a small rise,the detection accuracy of the improved model reaches 80.29%,which is7.95% higher than YOLOV3.Then,on Pascal VOC2007,the performance of the improved YOLOV3 is compared with other target detection algorithms.Compared with SSD300,Faster-RCNN and YOLO series models,the accuracy advantage of the improved YOLOV3 model is still obvious.