Research On Ship Target Detection And Localization Method Based On Deep Learning

Ships are important transportation carriers and military targets at sea.Accurate positioning and identification of ships can help dredge waterways and reduce water traffic accidents,and it is also of great significance for intelligent research of ship detection and construction of marine transportation industry.With the deepening of data and the continuous improvement of computing speed,deep learning has been rapidly developed in various fields of target detection,and the use of deep learning methods in the direction of ship target detection can not only enrich the achievements in the field of ship research,but also have practical significance for the intelligence of water transportation.In this paper,the following research is carried out on ship target detection recognition based on deep learning methods.1.Learn and research the current mainstream target detection and ship target detection methods,and conduct in-depth analysis of the deep learning technology in the current target detection field,and conduct experimental research based on official data sets.In order to better verify the detection performance of the method in this paper under various weather conditions,image enhancement technology is used to expand the dataset to improve the generalization ability and robustness of the model.2.For the feature extraction of Faster RCNN model,the influence of multi-level feature fusion on target detection is ignored.In the feature extraction stage,a bottom-up side fusion path network is specially added to strengthen the fusion of low-level feature information and high-level information.The experimental results show that in different weather environments,the method in this paper has a great improvement compared with other methods,and can better cope with different weather conditions.3.In the research on ship target positioning,it is found that the loss calculation is based on the comparison result of the candidate frame and the target frame as the basis for the regression of the bounding box,which lacks the estimation of the positioning uncertainty of the prediction frame and the target frame,resulting in the target positioning.deviation.This paper conducts in-depth research on the target detection and localization method,modifies the coding form of the prediction frame,redesigns the localization loss function using Gaussian distribution,and uses the variance voting method to improve the NMS localization effect.The experimental results show that the optimized detection model can not only improve the detection It can improve the accuracy of filtering out duplicate boxes and make the positioning of bounding boxes more accurate.It is beneficial to improve the situation of missed detection and false detection,and has good practicability in the application of ship detection in complex environment.