Few-Shot Compound Jamming Recognition Network Based on Image-Semantics Matching Mechanism

In recent years, compound jamming has introduced new challenges to jamming recognition task. Compared to Single Jamming Recognition (SJR), Compound Jamming Recogntion (CJR) exhibits greater variability in types, and the processes of data collection and labeling are more complex. However, few studies have focused on the few-shot CJR task. To address this gap, we propose an end-to-end Few-shot Compound jamming recognition network based on the Image-Semantic Matching mechanism (FC-ISM). We reformulates the jamming recognition problem as a matching problem between “jamming features" and “jamming types". The proposed method processes both single and compound jamming using a unified framework, so that there is no essential difference in the recognition logic between SJR and CJR in our network. For each specific jamming type, the network can be trained to establish the matching relationship using the single jamming dataset. When recognizing compound jamming, the pre-learned matching relationships can be directly transferred, requiring only a few samples to achieve high recognition accuracy. Specifically, first, we extract image features from time-frequency diagrams and semantic features from label words to characterize the jamming features and types. These feature vectors are then used to construct an image-semantic matching (ISM) graph by connecting each image feature node to all semantic feature nodes, modeling the full range of potential ISM relationships. Finally, we leverage a lightweight graph network block to excavate the optimal ISM relationship. Simulation results demonstrate that the average recognition accuracy of FC-ISM exceeds 95% with a training set comprising only 5 compound jamming samples. Compared to advanced recognition methods, FC-ISM not only achieves the superior recognition performance but also reduces computational complexity by approximately 90%, better aligning with the requirements of practical applications.