Research On SAR Image Target Recognition Algorithm Based On Multi-task Learning

Synthetic Aperture Radar(SAR)has the ability to perform remote sensing tasks in all-weather and all-day.By coherent accumulation of multiple echoes of the target,two-dimensional SAR images of the target can be obtained.In addition,microwave remote sensing has the penetration capacity of ground and vegetation,which is conducive to the detection of man-made building targets such as airports,ports,bridges,roads and military targets such as aircraft,tanks and ships.Therefore,it has important military and civil application value.In order to improve the recognition rate of targets in SAR images,the key steps such as preprocessing,feature extraction and classifier design need to be designed pertinently.This paper studies the design of the above key steps,and the main research results are as follows:Aiming at the shortage of training samples often encountered in SAR image target recognition,which leads to insufficient training of single-task learning classifier,a SAR target recognition method based on multi-task learning is designed.The method uses the split method of One vs.Rest(OvR)to disassemble the multi-class classification task into binary classification tasks,and then train the multi-task learning model.The experimental results show that the recognition accuracy of SAR image target recognition method based on multi-task learning is higher than that of single-task learning method.When using 40%training samples to train the classifier,the recognition accuracy of SAR image target recognition method based on Multi-task Relationship Learning(MTRL)is only 0.44%lower than that of using all training samples.The classification ability of the proposed method when the SAR image samples are insufficient is proved.Aiming at the problem that the time complexity of SAR image target recognition method based on multi-task learning is high and the distribution of positive and negative samples in each task data set is unbalanced and affects the classification effect,a strategy combining downsampling method and multi-task learning is proposed.The strategy uses the downsampling method to downsample the training data of each task in multi-task learning,and use the sampled data to train the classifier.Downsampling can change the imbalance of data distribution,and improve the generalization performance of the algorithm while reducing the running time of the classification algorithm.K-means,hierarchical clustering and random sampling are applied to the MTRL method and Clustered Multi-task Learning(CMTL)method respectively.The results show that the training time of MTRL and CMTL methods are reduced by 92%and 69%respectively after sampling the training data of each task.Compared with the other two sampling methods,when K-means is used as the downsampling method,both MTRL and CMTL methods obtain the highest recognition accuracy.It is proved that the proposed strategy can improve the application value of SAR image target recognition method based on multi-task learning.