Research On Intelligent Detection Of Ship Targets On The Sea Surface Based On Improved YOLO Algorithm

The perception and acquisition of information around the sea is one of the core technologies for the autonomous and safe navigation of intelligent ships.The most important prerequisite for the realization of intelligent autonomous avoidance and dynamic path planning for other ships is the real-time detection and high-precision identification of other ships within the visible range of the ship,which is of great significance to promote the unmanned and intelligent ship.At present,ship target detection algorithms based on deep learning mainly include: RCNN,Fast RCNN,FasterRCNN,SSD and YOLO algorithm,among which the deep learning algorithm with fast detection speed and high accuracy is YOLO algorithm.However,YOLO algorithm still has some shortcomings in target location,small target detection and detection of occluded ships,which greatly affects the accuracy and recall rate of ship target detection.In view of the shortcomings of the above-mentioned YOLO algorithm,the innovation points of this paper are as follows: 1.Aiming at the problem of inaccurate target positioning,this paper uses GIOU instead of MSE to achieve more accurate coordinate positioning;2.Aiming at the problem of small target detection difficulty,this paper combines the DSSD network to increase the acquisition of small target information to improve the accuracy of small target detection;3.Aiming at the problem of difficult detection of occluded ships In this paper,an improved NMS algorithm is proposed to detect the occluded ship.In order to show the effectiveness of the proposed algorithm,firstly,the traditional median filter,mean filter and Wiener filter are used to filter and denoise the image.In order to reduce the noise of the image to be identified,secondly,the ship training set is expanded by using random turning,cutting and other data augmentation methods.Finally,the network algorithm is compared with the current popular FPN network,SSD network,RetinaNet network architecture in terms of accuracy,speed and recall rate.The results show that the proposed algorithm can solve the problems of inaccurate ship target location,small target detection difficulties and ship occlusion,and the proposed algorithm is better than other advanced target detection algorithms in accuracy and recall rate,which effectively proves the effectiveness of the proposed algorithm.