Radar Signal Recognition Based on Transfer Learning and Feature Fusion

This study proposes a system for the automatic recognition of radar waveforms. This system mainly uses the obvious difference in Choi–Williams distribution (CWD) images of different modulated signals. We successfully convert problems related to radar signal recognition into problems related to image recognition. The classification system uses CWD time–frequency analysis of the detected radar signal to obtain its CWD image, which can be recognized by deep neural networks. To verify this method, a database containing 1800 images and 8 types of radar signal CWD images is established. Although a convolutional neural network exhibits strong expression, it is not suitable for training a small-scale database. To solve this inadequacy, an image classification algorithm based on transfer learning and design experiments is proposed. This algorithm is intended to fine-tune three different pre-training models. This study also integrates the texture features of the image with the depth features extracted using the depth neural network to compensate for the shortcomings of the depth features in expressing image information. The simulation results indicate that the method can still be used to effectively recognize radar signals at a low SNR.