Study On Ship Detection Method Based On Deep Learning

As important carriers and military targets on the sea,accurate positioning and identification of ships is conducive to improving navigation safety and work efficiency,and has important practical and strategic significance for monitoring maritime traffic,safeguarding national maritime rights and interests and maritime security.In the past10 years,people have done a lot of work on automatically extract targets from satellite images,and achieved some remarkable results,but there are still many shortcomings in practical applications.With the development of deep learning,target detection technology has made breakthrough progress in many fields.Therefore,the use of deep learning-based target detection methods for ship detection research has practical feasibility and important academic significance.In this paper,based on the deep learning method,the following studies are carried out on ship detection and identification:1.The mainstream ship detection algorithm and target detection algorithm are extensively investigated,and the detection ideas and innovation of each method are analyzed and studied.Then,the data set of ship detection is analyzed,and training data set is made,meanwhile,various methods were used to enhance the data.Experiments show that using data enhancement can effectively improve the generalization ability of the model.2.In this paper,we use the RetinaNet model to detect the ship target,and build the RetinaNet detection network.The experimental results show that the performance of the original RetinaNet for ship detection not very good and needs to be improved.On the basis of the original RetinaShip network,the SSH module is added to enhance the feature extraction.By adjusting the proportion of positive and negative samples,the influence of negative samples is reduced.The anchor box obtained by clustering is used to make the positioning of prediction box more accurate.Then,the model RetinaShip is proposed for ship target detection.Experiments show that RetinaShip can effectively improve the precision of ship target detection,and the average precision can reach93.28%.3.The popular YOLOv3 algorithm has been deeply studied and improved.The cluster algorithm is used to redesign the anchor box which accords with the ship data set,and the loss function is improved by using the CIoU.The results show that the optimized detection model can obviously improve the detection precision and make the AP of the detection reach more than 93%.This paper also implements the new YOLOv4 algorithm,and the average precision of ship target detection reaches 95%.Then,a lighter mobile net is used to replace the backbone network of the YOLOv4 model,which reduces the amount of parameters and improves the detection speed on the premise of ensuring the accuracy;a lightweight YOLOv4-Lite network is also realized,which has less parameters and faster detection speed,and can be used in some scenes with high real-time requirements.