Research On Synthetic Aperture Radar Target Recognition Method Based On Small Sample Set Condition

Synthetic aperture radar(SAR)technology can transform complex automatic target recognition image data into useful intelligence information through image processing and information interpretation,which is one of the key technologies to realize the military application of SAR imaging system,and has a wide range of application requirements in intelligence reconnaissance,precision strike and other fields.The military targets,especially high-value strategic targets,are often characterized by high mobility,strong concealment and randomness.Therefore,the corresponding target sample sets show the characteristics of small sample set,such as insufficient data within category and imbalanced data between categories.However,the data-driven ATR model usually shows a strong dependence on the data scale and data distribution of the SAR target sample,which forms a contradiction between the data supply of the SAR target sample and the data demand of the ATR model.The target recognition task with small sample set has also become a key problem in SAR ATR technology in recent years.In order to break the strict constraints of existing recognition theories on the data scale and data distribution,an automatic target recognition method with knowledge expansion,knowledge screening and knowledge reasoning is proposed under the premise of small sample set.Combined with the characteristics of small sample set in the SAR ATR task,the thesis condensed several scientific problems that need to be solved urgently,and conducted in-depth research on the SAR ATR based on small sample set condition.The research content and main contributions of the thesis include the following aspects:(1)To solve the problem of high quality directional generation of SAR target samples under the condition of insufficient data within category,a label directed SAR target sample generation method is proposed.Based on the reconstruction mechanism of generated sample data distribution,the proposed method uses effective distance measurement to describe the distance between real sample distribution and generated sample distribution.Then,the proposed method establishes the mapping between the label coded information and the generated sample category,and realizes the high quality directional expansion of the corresponding discriminant information under the condition of insufficient data within category.(2)Aiming at the problem of effective utilization of generated counterfactual SAR target samples under the condition of insufficient data within category,a SAR target sample recognition method based on counterfactual sample filtering is proposed.Based on the premise of a large number of generated counterfactual SAR target samples,the consistency cognition between different weak ATR models is established by using the significant differences between different batches of generated samples,so as to filter out the counterfactual SAR target samples conducived to the ATR model.It avoids the performance degeneration of ATR model caused by counterfactual sample generation,and realizes the significant improvement of ATR model recognition performance under the condition of insufficient data within category.(3)In order to solve the problem of demand allocation of generation model under the condition of imbalanced data between categories,a demand-driven SAR target sample generation method is proposed.Starting from the inherent limitations of the generation model under the condition of imbalanced data,the proposed method avoids the risk of quality degradation of generated SAR target samples by executing different decision logic in layers of the generation model.It establishes the mapping between various target samples and the corresponding generation demands,which improves the attention of the generation model to the minority category SAR target samples.The proposed method realizes the effective expansion of the discrimination information of the minority category SAR target samples in the imbalanced dataset.(4)Aiming at the cost-sensitive awareness construction problem of ATR model under the condition of imbalanced data between categories,a cost-sensitive awareness-based SAR automatic target recognition is proposed.Under the premise of supplementing the discriminative information of minority category SAR target samples,the proposed method establishes the mapping between various recognition results and corresponding recognition cost,so that the ATR model can autonomously learn the recognition cost of various target samples from the imbalanced dataset.It constrains the empirical preference of the ATR model for the majority category SAR target samples,which achieves stable improvement in the recognition performance of the ATR model under the condition of unbalanced data.Compared with the existing methods,the proposed method can significantly improve the SAR target recognition performance under the condition of small sample set.It has a certain role in promoting the application of SAR target recognition technology in the case of limited data acquisition.