Target Recognition Based On Multi-task Learning In SAR Imagery

Synthetic aperture radar(SAR)is a sensor which carrying out ground observation actively with characteristic of all-weather,day/night.It is vital to achieve target recognition in SAR imagery while the main obstacle is labeled sample is not sufficient due to sensitivity of SAR imaging with pose variation.Multi-task learning(MTL)is a mechanism where related models trained simultaneously based on heterogeneous feature or source,which is beneficial for target recognition performance improvement.Purpose for resolving labeled sample deficiency,paper pays attention on three key problems based on MTL with SAR imagery multiscale analysis: feature selection,sparse representation and sparse approximation algorithm:(1)A feature selection method based on sparse vector distribution similarity with SAR imagery multiscale analysis is proposed to meet requirement underlying MTL framework.Firstly,sparsity distribution of coefficient from sparse representation on validation set is calculated.Then relatedness between two task is measured by sparsity distribution.Finally,correlation information entropy derived by similarity matrix is utilized to select appropriate feature set.(2)On purpose of resolving ineffectiveness of sparse representation resulting from sample deficiency,a dictionary learning method based on linear local constrained sparse representation under multi-scale feature is proposed.Based on MTL,dictionary optimization with linear local constrained can achieve target recognition improvement when samples are limited.The experimental results verify that the proposed method can improve the performance of SAR target recognition with imagery.(3)A matching pursuit algorithms with weighted neighborhood is studied.Under the framework of MTL,the multi-scale sparse vector of the residual is weighted by the neighborhood,and the atom is selected to realize the matching pursuit.The target classification is realized with the multi-scale reconstruction error generated from approximation signal of scale.Experimental results show the effectiveness of proposed algorithm.