Research On Deep Transfer Learning For Radar Target Recognition

The data-driven deep learning method has been used as a radar automatic target recognition research due to its powerful feature extraction capability.Deep learning can automatically learn the features of data from a large number of labeled training samples.In contrast to traditional machine learning methods,deep learning often extracts more semantic features.However,there are usually less radar data with labels,which cannot meet the requirements of deep learning,resulting in serious over-fitting.Moreover,the radar sensor is susceptible to noise interference,and the radar echo is penetrated by the relative angular displacement between the radar and the target,and the deep learning is usually reset in the multi-scene target recognition task.Here,the shortcomings for deep learning for radar target recognition will be proposed to propose transfer learning solutions.First of all,this paper starts from the transfer learning of optical images to radar target images,and introduces the basic principles of deep learning and transfer learning.And using three different neural network structures to test the feasibility of optical image to radar image migration.The results show that transfer learning can indeed improve the convergence speed compared to the randomly initialized model.However,the improvement of recognition accuracy is not obvious,and even causes negative transfer,leading to a decrease in accuracy.Then,this paper studies the transfer learning from unlabeled radar images to tagged radar images.Unsupervised feature learning is used to extract common features from unlabeled data,and common features are passed to target tasks through parameter sharing.The original network is then adapted to the target classification task by fine tuning techniques.In general,this paper combines transfer learning with semi-supervised learning to maximize the use of training data.It has been verified that the performance of the method here has been significantly improved with few labeled training samples.Finally,transfer learning between different diffracted radar images has not been studied here.In this paper,the different radar images are decomposed in the high-level feature distribution of the neural network by the method of adversarial learning,and can be discriminated by the same feature classifier.It is verified that the proposed method greatly improves the performance of deep learning in multi-level radar target recognition tasks.Moreover,the algorithm proposed in this paper also has strong anti-noise performance.