Research On Key Technologies Of Ship Target Detection And Formation Identification

Ship target detection and formation identification have always been the focus of maritime research in various countries.As the core part of ship detection system,ship target detection and formation identification play an important role in Marine ship management,maritime traffic and transportation management,ship rescue and rescue,and intention acquisition of ship formation.At present,the ship target detection has not been deeply studied,and it still remains the simple migration application of the general target detection algorithm,and lacks the idea of global optimization.The research and analysis of ship formation identification are relatively few,and it is difficult to apply in practice.In this context,this paper carries out research on the algorithm of ship target detection and formation identification,and innovatively links them into a whole for research.The innovation points and research results of this paper are as follows:(1)In general target detection algorithms,the rotation detection box is used to replace the horizontal detection box,which usually introduces noise and other interference information and affects the performance of the algorithm.In this paper,Ro I with Mask Filter(MRo I)is proposed to reduce the impact of noise on location and regression by adding spatial direction information to the feature and using Mask Filter.Experimental results show that MRo I network can improve the quality of feature semantics,and thus improve the algorithm performance.(2)The traditional regression loss function is used in the existing ship target detection algorithm,which is not helpful to the prediction of rotation angle.The distribution of Angle parameters in the rotation detection Box is different from the element distribution in the horizontal detection Box.Therefore,starting from the original definition of the Loss function and based on the regression prediction network parameter distribution,this paper designs the Kullback-Leibler Divergent divergence of the Rotate Bounding Box Loss(RKL Loss)suitable for angle element regression.The experimental results show that RKL Loss not only improves the recall rate,but also has gain effects on other performance indicators.(3)According to the position and bounding box information obtained from ship target detection,global feature descriptors and local feature descriptors are constructed for ship formation identification.Based on the construction of feature descriptor and considering the scarcity of training samples,a new similarity measure function is designed to reduce the dependence of formation recognition on training samples.Experimental results show that the proposed algorithm has high robustness and is suitable for practical application.