| Inverse synthetic aperture radar image recognition is widely used in civil and military fields.However,in actual radar target recognition applications,especially under non-cooperative target conditions,it is difficult to obtain enough training samples to meet the training requirements of the model,resulting in deep learning technology in small sample conditions.The performance of ISAR image recognition is not good.Current data enhancement,transfer learning,and meta-learning are important methods to solve the recognition under few sample conditions.Especially under the condition of very few samples,metric learning method in meta-learning shows excellent performance.Aiming at the problem of ISAR image recognition under small sample conditions,this paper studies the application of metric learning methods to small sample ISAR Like recognition.The main research work and results of this article include:1.The first chapter reviews the research progress of ISAR image recognition and measurement learning at home and abroad,analyzes the shortcomings of traditional methods,and introduces the experimental data used in this article.2.Chapter 2 introduces the principle of metric learning.Experiments on the ISAR simulation data set verify the effectiveness of the metric learning method in solving small-sample ISAR image recognition,and focus on the analysis of two limitations of the metric learning method.It is in metric learning that a small number of label samples are random,so each type of prototype extracted by the model is also random,which leads to the model performance is not robust enough.The test of the metric learning method is to randomly select multiple sampling tasks for classification and recognition,but each recognition task of the model is unknown,so the model cannot extract the most separable feature of each unknown task.Second,for ISAR image recognition,the measurement capabilities of existing measurement learning methods need to be improved.3.In Chapter 3,in view of the problem that the ISAR image recognition feature in metric learning does not reach the best separability.This paper proposes a Transformer based ISAR image recognition method for metric learning.This method introduces a Transformer composed of triple attention modules to adapt Learn a more separable metric space,thereby improving model recognition performance.Finally,electromagnetic calculations and darkroom measurement data simulation experiments are used to verify the effectiveness of the proposed model in small-sample ISAR image recognition.4.In Chapter 4,in view of the limitation that the existing metric learning methods are sensitive to the spatial position relationship of ISAR images,which leads to the poor measurement ability of the model,this paper proposes a dual attention relationship network for small-sample ISAR image recognition.The model first introduces split based convolutional operation(SPconv)to eliminate pattern redundancy in ISAR image features,extracts the effective features of ISAR images,and takes the average of the labeled sample features to obtain the category prototype,and then embeds a dual correlation attention module(DCA)To perceive the positional relationship of ISAR images and compare related semantic objects or fine-grained features at different locations.Finally,a learnable non-linear classifier is used to measure the similarity between the test sample and each type of prototype(category center)for classification.The effectiveness of the proposed model in small-sample ISAR image recognition is verified by electromagnetic calculation and darkroom measurement data simulation experiment..5.Chapter 5 summarizes the research work and innovation content of the full text,and points out the future research directions that are expected to improve the performance of the model. |
