Research On Ship Target Detection On Water Video Method Based On Deep Learning

The rapid development of inland shipping and marine shipping has promoted economic growth,but it has also caused more and more accidents such as ship collisions and ship reefs.Illegal fishing and parking of illegal ships also frequently occur.Therefore,it is of great significance to realize the automatic detection of the target of the ship on the water for the management of the ship on the surface.Firstly,for the complex water surface environment,the movement of the ships,the camera shake and other factors cause the quality of the ship's image to decline,this article preprocesses the ship's video image.After comparing with traditional image deblurring algorithms,this paper adopts the Parameter Selective Sharing and Nested Skip Connections(PSS-NSC)method based on deep learning with better image deblurring effect to further improve the quality of ship images.Due to the lack of open source ship data sets,this paper refers to the commonly used open source Pascal VOC data set format to make ship data sets for the subsequent use of ship target detection network models.Secondly,using the ship data set produced in this paper,Faster Region-based Convolutional Neural Network(Faster R-CNN),Single Shot Multi Box Detector(SSD)and You Only Look Once v3(YOLOv3)was trained and tested.The performance indexes such as accuracy and detection speed of the three algorithms are compared,and then the YOLOv3 network model with both speed and precision is selected for subsequent research on ship target detection.Then,this paper proposes an improved YOLOv3 algorithm,which uses Darknet-53 as the feature extraction network.The improved algorithm innovatively incorporates the salient region features obtained by the Frequency-Tuned(FT)Salient Region algorithm.The obtained distinctive regional features are processed by the CBL(Conv+BN+Leaky Re LU)component network and connected with the deep features extracted by the Darknet-53 network model to the Batch Normalization(BN)layer.The Concat method is used for feature fusion,and finally the convolution operation is performed by 3?3 convolution kernel to obtain the final feature map,which makes the network more powerful in description and robustness,and achieve higher target discrimination;Soft Non-Maximum Suppression(SoftNMS)method is used to replace the original Non-Maximum Suppression(NMS)method of YOLOv3 algorithm,which reduces the risk of mistaken deletion of frames,missed detection and false detection of targets,thereby improving the detection rate of the model.Experimental results show that the improved YOLOv3 algorithm improves the accuracy of video ship target detection,optimizes the performance of video ship target detection,and further improves the detection effect of small targets and overlapping targets.Finally,a video-based ship target detection system on water was designed using the Microsoft Foundation Classes(MFC)framework.The overall performance test of the system showed that the system can accurately and real-time detect complex surface ship video.