Few-shot ISAR Target Recognition Method Based on Improved Prototype Networks

Inverse Synthetic Aperture Radar (ISAR) possesses unique advantages such as all-weather capability, long-range coverage, and high resolution, making it an important player in target perception and air defense early warning. However, when performing ISAR image recognition, classical deep learning algorithms suffer from reduced recognition accuracy and poor stability with fewer training samples. To improve classification accuracy and stability, we proposed a few-shot ISAR images target recognition method based on improved prototypical networks (IPN). Unlike the algorithm before improvement, combined with the idea of skip connections between modules in residual networks, we built deep convolutional neural networks feature extraction module called ResCNN for deep extraction of sample features, and at the same time, avoiding the decrease in the stability of the model caused by the gradient vanishing and gradient explosion during the training of deep neural networks. After extracting the sample features and mapping them into embedding vectors based on the ResCNN module, we compute the prototypical space based on the embedding vectors of the samples of the support-set, and realize the prediction of the target category by calculating the Euclidean distance between the embedding vectors of the samples of the query-set and the various types of samples in the prototypical space. Finally, in order to verify the performance of the proposed method, we conducted simulation experiments for plane classification and analyzed the experimental results for comparison. The recognition results of the measured data demonstrate that this method achieves higher average recognition accuracy and lower standard deviation under few-shot conditions. Therefore, the proposed algorithm ensures the stability of the model with improved accuracy and can be applied in few-shot ISAR image recognition tasks.