Multi-Task Learning For Synthetic Aperture Radar Target Recognition

Since the 1990 s,synthetic aperture radar(SAR)target recognition technology(ATR)has been a forward topic of SAR,which can effectively penetrate the cover regardless of the weather condition.In recent years,SAR ATR has made progress in image preprocessing,feature extraction,feature fusion and target classification.The traditional SAR ATR algorithms are based on single task learning framework.It is vital to extract the relevant information due to sensitivity with pose variation and non-linear features of SAR imaging.Paper pays attention on SAR ATR algorithm based on MTL which is beneficial for recognition performance.The main work includes the following sections:1.The characteristics of SAR images are analyzed,and the key techniques and difficulties in SAR ATR are studied.The traditional KNN and SVM are deeply analyzed.Multi-task learning(MTL)is introduced to reduce the impact of sensitivity with pose variation.2.A SAR ATR method based on RMTL algorithm is presented,which is based on the MTL framework.The relevant information are simultaneously learned and fully extracted among multiple targets to reduce the impact of SAR image attitude sensitivity.The experimental results indicate the superiority of RMTL algorithm.In order to avoid the influence of outline tasks,a SAR ATR method based on MTRL algorithm is proposed.In view of the limitations of MTRL algorithm in SAR ATR,CMTL algorithm is proposed to be applied.CMTL algorithm can accurately measure the relatedness among multiple targets.Experimental results and comparisons with several competing algorithms corroborate the efficiency and effectiveness of the CMTL algorithm.3.K-CMTL algorithm for SAR target recognition is proposed.In this method,the latent grouping structures of CMTL algorithm is used.In view of the non-linear features of SAR imaging,the clustering structure in the original space is extended to the non-linear space,and the relationship between multiple targets in the non-linear space is constrained.Instead of performing model decomposition to cope with various structure elements,a regularization term to formulate an unconstrained and non-smooth convex optimization problem are directly imposed.The accelerated Proximal Gradient(APG)method is applied to derive efficient solutions.4.The K-CMTL algorithm is simulated based on the MSTAR.Firstly,the parameters are analyzed under the standard operating conditions(SOC),and the comparison experiments between single-task and multi-task are given.Secondly,K-CMTL algorithm is tested under the SOC,the Extended Operating Conditions(EOC-I)and the Extended Operating Conditions(EOC-II)respectively.Comparing with the existing methods,K-CMTL achieves better results in empirical study.