Research On Ship Target Tracking And Recognition Technology In Optical Remote Sensing Images

Remote sensing tracking and recognition is an important means to master ship target dynamics,obtain military information and safeguard maritime rights and interests.It plays an irreplaceable role in national defense construction and economic development and has important strategic significance and application value.The development of remote sensing technology provides a solid data guarantee for maritime reconnaissance and surveillance.It is the core of tracking and identification task to obtain the position,dynamic and category information of ship targets based on remote sensing data.Since the remote sensing platform is a top-down imaging perspective,ship targets show arbitrary orientation and dense array,and contain many background regions,which brings great challenges to accurate target location and dynamic tracking.The similarity between classes and the differences within classes caused by various imaging factors increase the difficulty of fine-grained classification of ship targets.Therefore,how to achieve accurate,stable and efficient ship target tracking and recognition is a challenging task.In order to solve the above problems,this paper carries out research from three aspects:detection,tracking and fine-grained recognition,and explores automatic and intelligent ship target tracking and recognition methods.Specific research contents are as follows:（1）Aiming at the problems of arbitrary orientation,dense arrangement and extreme aspect ratio faced by ship target detection tasks,a directional frame detection algorithm based on dynamic adjusting labels strategy and gradient truncation mechanism is proposed to improve the target positioning accuracy.Based on the classification method of intensive coding angle prediction,the dynamic adjusting labels strategy is used to strengthen the coupling between code points,and to guide angle coding to carry out directional learning,so as to give full play to the accuracy advantage of long coding.Then,a truncation function is added for each code point loss through the gradient truncation mechanism to control the gradient return process to balance the difficulty of learning code points,prevent overfitting and underfitting phenomena,and improve the learning effect of coding.Experimental results on open data show that the proposed method can achieve high precision ship target location,and the detection accuracy reaches 90.50%.Compared with other methods,it has an improvement of 1.21%～20.03%,and the high precision index m AP₈₀ has an improvement of 7.63%compared with the baseline method.（2）Aiming at the problems of tracking interruption and target loss faced by ship target tracking missions,an inertial tracking model IPMOT based on TBD normal form is designed.Through inertial prediction of directional tracking box parameters,track completion can be achieved in case of missed detection by the detector,so as to avoid track interruption and improve tracking stability.Due to the angle mutation of the orientation frame at the boundary,the tracking frame will rotate sharply in this case,which seriously affects the target positioning accuracy.In this regard,IPMOT ensures the continuity of the input angle sequence by angle correction,and limits the output angle sequence within the main value interval to eliminate the influence of boundary discontinuity.The comparison results on the tracking data set show that many indexes of IPMOT are better than other methods,and MOTA,IDF1,MT,IDs and Frag are 61.2%,65.7%,48.8%,163 and 977,respectively.This indicates that IPMOT effectively improves tracking stability,reduces the number of track interruption and ID jump,and reduces the probability of target loss.（3）A feature asynchronous learning network based on contrast learning architecture is proposed to solve the difficulty of type recognition caused by inter-class similarity and intra-class difference in fine-grained ship target recognition tasks.The network obtains high quality local attention diagrams by decoupling local features of ship targets,and uses counterfactual causal attention structure to learn the logical relationship between prediction and input to eliminate false correlation and enhance the logic of prediction.By using the local attention obtained from decoupling branch learning,the main features are weighted to obtain fusion features,and then the feature reassociation and global feature aggregation are carried out to realize the clustering of feature vectors.Finally,the feature vector is input into the classifier to get the fine-grained recognition result.The results on two public data sets show that compared with other methods,the OA and AA of the proposed method are improved by 3.03%～10.18%and 3.36%～9.89%,and the recognition accuracy is better.