A CrossNet-Based Prototype Network with Similarity Guidance for Few-Shot Radar Target Detection

Few-shot radar object Detection poses a challenging task of recognizing the radar target from with only few numbers of annotated example as supervision. In this article, we propose Cross-SGNet, a Prototype Network based on CrossNet to tackle the few-shot radar target detection problem. We first adopt a masked average pooling strategy for achieving accurate extraction of target features by masking the background information in the support samples. This strategy enables us to obtain feature vectors of the target in a specific high-level feature space. Additionally, we leverage a similarity guidance strategy to measure the cosine similarity between the query feature and the support vectors by location. These similarity maps can guide the process of detecting radar targets. Meanwhile, to improve the accuracy of target size predict results, we propose an efficient method that employ the shape feature in Heatmap to extract the width and height information of the radar target. Furthermore, our Cross-SGNet fully exploits knowledge from a limited amount of information and provides better generalization on few-shot radar object detection. We conduct experiments on a few samples of radar data to compare CrossNet and Cross-SGNet. Significantly, our model achieves the precision score of 96.7% and recall score of 74.0%, surpassing CrossNet model. These results demonstrate that Cross-SGNet has the potential to improve the detection performance of radar systems with limited annotated examples.