Research On Incremental Learning Methods For Automatic Target Recognition

Automatic target recognition(ATR)is a target recognition technology that is dependent on the data acquired by the sensors.It is able to judge the target type or attribute through pattern recognition and machine learning algorithm.ATR is one of the key technologies to improve the information sensing ability and realize the application of sensor technology.So it is valuable to research deeply.With the rapidly developing and wide applying of sensor technology,massive data has been accumulated,and the amount of data is still increasing.Increment data plays a positive role in the improvement of target recognition ability and the expansion of application scope.Meanwhile,effective processing and efficient utilization of incremental data has also become a key issue in target recognition in recent years.This needs the recognition algorithm to learn new training samples and new target types.In this process,ATR should gradually adapt to the changes in sensor deployment environment and data distribution.In final,the ATR system acquires stronger recognition ability and completes more complex and diversified recognition tasks.To improve the ATR methods on the ability of continuous learning and updating knowledge,this work analyzes the incremental data characteristics and the main scientific problems of military target recognition,and then researches incremental learning methods for ATR.The main research contents and contributions include the following aspects:(1)In order to solve the problem of updating recognition models rapidly,a generalized-sparse-constraints-based incremental non-negative matrix factorization method is proposed.This method starts from disassembling the objective function and only calculates the new part of the objective function corresponding to the new sample.It is able to avoid repeating training existing samples when updating the model.Considering the sparse characteristics of the target,the generalized norm-constrained objective function related to incremental samples is established.Thus,the accuracy of the new sample calculation and the accurate performance of the updated recognition model are improved.(2)Aiming at the problem of significant training sample selection in increment data processing,a class-boundary-exemplar-selection-based incremental learning method is proposed to integrate the key knowledge of existing data.Based on the local statistical and geometric information of the data distribution,the method selects boundary exemplars in the overlap,edge,and interior regions between classes respectively by using the herding effect.When new classes are updated,the previous training data distribution is reconstructed to update the boundary exemplar set.The proposed boundary exemplar method is able to improve the recognition performance on the classifiers under a fixed number of exemplars.(3)To solve the problem of new unknown target recognition,an inter-class-extremum-distance-based open set incremental learning method is proposed.Starting from the risk function of open set recognition,the inclusion probability decay model which can quantify the occurrence probability of new classes is studied.In order to balance the performance of the known classes' classification and the new unknown classes' rejection,an extreme-value-theory-based open set classification model is established relying on the distance range of the existing classes' distribution.Finally,combined with the edge exemplar selection method,only the inter-class extremum distances are used in the modeling to realize incremental learning.The proposed open set incremental classifier is able to complete the classification of training classes and the rejection of new classes at the same time and realize the local update of the recognition model.(4)An incremental learning method based on distribution reliability assessment is proposed to evaluate the updating benefit of new unlabeled samples.This method considers the distribution expansion mechanism of incremental data.It evaluates the“in-of-distribution” reliability of new samples by calculating the local density difference and evaluates the classification reliability of predicted samples by calculating the inter-class overlap distance ratio.New samples with low “in-of-distribution” reliability are used to create new classes,and new samples with low classification reliability should be assigned new correct labels manually.The updating mechanism based on reliability assessment can reflect the updating value of new samples,which could reduce the cost of manual annotation and improve the updating benefit.In this research,a data-efficient utilization mechanism based on incremental learning is introduced for automatic target recognition.The theories and methods of target recognition are enriched in the aspects of incremental feature extraction,exemplar management,open set recognition,predictive reliability assessment,etc.This work could provide technical support for the construction of an automatic target recognition system with autonomous learning ability.