Potential Ability Analysis for SAR Target Recognition with Small Samples

‍ ‍With the encouraging results of deep learning in the field of computer vision， it is possible to realize recognition of time-sensitive targets in SAR images based on deep learning technology. The recognition of time-sensitive targets with SAR images has attracted a lot of attentions. Limited by objective conditions， the cost of high-quality SAR target samples acquisition is high. In addition， SAR system is very sensitive to imaging parameters and target attitude， which makes the target recognition under the condition of small samples more difficult. The target recognition potential of MSTAR dataset is deeply excavated in this paper， and the performance potential of vehicle target classification and recognition of 10 types of SAR images is fully studied and analyzed. To measure the target recognition potential of SAR under the condition of different number of samples， and at the same time reduce the randomness of selecting target samples， using different numbers of measured training samples to generate full-angle training data sets is proposed in this paper to standardize the samples participating in the training. The expanded training sample set is used as the standard template data set. By traversing the template data set， the likelihood ratio similarity measure （Likelihood Ratio Similarity Measure， LiRSM） is used to measure the target similarity. By using the gray statistical characteristics of SAR image， the change detection is constructed based on change detection technology， and realized the classification and recognition of SAR vehicle targets. Based on MSTAR dataset， the research on 10-way-N-shot target recognition is carried out by using different numbers of real SAR training samples， and the basic benchmark test is formed through experimental comparison， which is convenient for other scholars to further carry out comparative analysis of target recognition under the condition of small samples.