Abstract A key challenge in ensuring the quality of high-sensitivity observations from the Chinese Sky Eye FAST is the efficient suppression of radio frequency interference (RFI). However, the traditional thresholding method cannot adapt to the complex baseline noise because it relies on manual parameter adjustment and assumes a constant threshold value, it also faces bottlenecks such as the failure to detect weak interferences and the ambiguous positioning of the boundary. This paper proposes a deep learning framework (WF-SwinUnet) based on the multiwindow fusion Swin Transformer, which achieves end-to-end accurate RFI segmentation through the collaborative design of an adaptive multiscale window mechanism and a bottleneck residual network. Experiments show that compared with the SumThreshold method, this model improves the accuracy, recall and F1 score by 23.4%, 32.1%, and 30.1%, respectively, improves the processing efficiency by 60% (reducing the time from 10,440.95 to 4387.87 ms), and significantly improves the accuracy of boundary positioning, reducing the boundary error (HD95) by 73.6% compared with ArPLS-ST. For weak narrowband RFI (signal strength close to background noise), which is difficult to detect with traditional methods, the model F1 score reaches 0.6808, avoiding signal loss due to threshold overload. In addition, among the six FAST observation targets, the model’s average F1 score reached 0.818, an improvement of 8.3% over the existing best deep learning model (RFI-Net). Ablation experiments verified the synergistic effect of the multiwindow fusion mechanism and the residual module, which together improved the overall performance by 21%.
Li et al. (Thu,) studied this question.