Research On Target Recognition Of SAR Images Based On Representation Learning

Synthetic Aperture Radar(SAR),as a representative of microwave remote sensing,has become an important part of China's Earth observation system based on its high-resolution imaging characteristics.Compared with remote sensing technologies such as visible light and infrared,SAR work is not affected by illumination and climatic conditions.It has all-day and all-weather working characteristics and has certain penetrating power.The target recognition technology based on SAR image has been widely used in military and civilian fields,and has attracted the attention of scholars all over the world.Existing SAR target recognition methods based on sparse representation,collaborative representation,etc.,have problems such as weak generalization performance,insufficient robustness,and poor performance in configuration recognition.In response to these problems,research is carried out,and the main tasks are as follows:(1)Aiming at the typical problems of existing sparse representation SAR target recognition algorithm,the generalization performance is insufficient,and there is a lot of redundant information and interference information in the dictionary,the SAR image target recognition algorithm based on two-stage multitask representation learning is proposed.The algorithm first extracts three types of features of SAR targets,and recognizes the recognition under single feature as a classification task,thus modeling multi-feature recognition as multitask learning.The algorithm is divided into two stages of multitask learning.On the basis of analyzing the target characteristics of SAR,the first stage uses multitask sparse learning to obtain the local atom set of the current test sample in the global dictionary,and uses the local atom set as the dictionary of the second stage multitask learning.The second stage uses multitask collaborative learning.The advantage of this algorithm is that it integrates the multi-feature recognition ability of target recognition through multitask learning,which improves the robustness and generalization ability of the algorithm.At the same time,the two-stage multitask learning mode is used to eliminate the interference of a large number of unrelated dictionary atoms,thus improving the recognition accuracy.The target recognition experiment is carried out by using the publicly released SAR database.The experimental results show that the proposed algorithm has higher recognition accuracy and better robustness than the existing sparse representation SAR target recognition algorithm.(2)Aiming at the low accuracy of the existing SAR target recognition algorithm for configuration recognition and not fully exploiting the low-dimensional structure of SAR target image data,a SAR target configuration recognition algorithm based on multi-feature low-rank representation fusion is proposed.The algorithm first extracts three types of features of the SAR target.Then,based on the low-rank representation model of classwise block diagonal structuring,a low rank representation under each class of features is established.Then,the results of the low-rank representation of the three types of features are fused.This fusion combines the two mechanisms of local azimuth fusion and Bayesian fusion,which can further improve the discriminating ability between different target configurations.Based on the different configuration target data in the publicly released SAR database,the configuration recognition experiment is carried out.The experimental results show that the algorithm obtains better configuration recognition accuracy and is superior to the existing algorithms.