Research On Unmanned Ship Target Detection Method Based On Deep Learning

In recent years,the intelligent ship technology has been continuously developed,and the strategic position of the ocean is also increasing.In order to improve the supervision and control capabilities of the ocean,countries have gradually increased their research on unmanned boats on the surface.In order to realize its autonomous obstacle avoidance function when an unmanned boat is sailing in a complex sea environment,it is necessary to sense,recognize and calibrate the ship target on the sea in real time.Taking this as an entry point,this dissertation uses deep learning algorithms to study the problem of ship target detection,focusing on the needs of ship detection tasks in the course of unmanned ship navigation,focusing on improving the accuracy and speed of detection algorithms.The main research contents of this article are as follows.In view of the lack of sea surface ship data sets,this dissertation built a ship image data set.A large number of ship pictures are crawled through the web crawler,and the ship pictures are manually marked.The data set contains more than 6,000 pictures,including five types of ships,cargo ships,passenger ships,sailing ships,warships and ordinary ships.Laid a good foundation for training and evaluation.At the same time,the established data set makes up for the vacancy of the existing ship data set,and provides effective resources for future ship inspection tasks.This dissertation establishes an improved Faster R-CNN ship target detection deep learning algorithm.Use the Faster R-CNN algorithm to train on the self-built training set to complete the ship detection task.At the same time,in order to further improve the detection accuracy of the network,the basic feature extraction network is improved,and a deeper residual network is selected to replace the original VGG.-16 network,and use clustering algorithm to redefine the anchor box in the network to make the anchor box more targeted.On this basis,the soft-nms algorithm is used to replace the original anchor box screening algorithm,to a certain extent,reducing the problem of missed detection in the case of high occlusion.Experiments verified that the network mAP value using the improved algorithm increased from 81.97% to 86.04%,which verified the effectiveness of the optimization algorithm.Aiming at the problem of slower detection speed of Faster R-CNN network,thisdissertation further studies the ship detection algorithm based on R-FCN,and gives the optimization strategy of the algorithm.During the research,it was found that the RoI-wise sub-network in Faster R-CNN algorithm needs to be calculated separately,which slows down the network detection time.Detection speed;at the same time,in order to solve the position insensitivity problem of the deep network,a position-sensitive score map is used.In order to further optimize the accuracy of the R-FCN detection network,in addition to the improved anchor box and soft-nms algorithm,the OHEM algorithm and Modify the classification loss to focal loss to solve the problem of imbalanced sample categories in the network,and further improve the accuracy of network detection.This article uses the established ship data set to compare the mAP value and detection accuracy of the algorithm before and after optimization under the Ubuntu system.The results show that the improved R-FCN network is faster than the improved Faster R-CNN network in running speed and The accuracy has been significantly improved,which can meet the requirements of unmanned ships to detect sea ships in real time.