Few-shot learning for signal detection in wideband spectrograms

Signal detection in the wideband plays an important role in spectrum sensing or reconnaissance tasks. Considering the visualization benefits of the spectrogram and the great developments in deep learning object detection, an increasing number of researchers have implemented deep learning-based signal detection in wideband spectrograms, which obtained remarkable performance. Most existing detection models rely on the availability of abundant labeled training data, but for signal classes with little labeled training data, the detection performance will deteriorate significantly. In this paper, a few-shot signal detection model is proposed to solve this problem. The proposed model is pretrained on abundantly labeled base signal classes and aims to detect novel classes given only a few labeled samples. The model is built on an existing base detector designed specifically for signal detection, and a class-specific convolution kernel generator (CCKG) is proposed to generate convolution kernels by template signals for predictions of signal center frequency and shape attributes. Benefiting from a three-stage meta-learning procedure, the CCKG can play a significant role with only a few input samples. Comprehensive experiments with a simulated signal superimposed on real background dataset and a real-world dataset demonstrate that the proposed method yields significantly better performance than the well-established baseline models.