Abstract The increasing sensitivity and data volume of modern radio telescopes mean that mitigating radio frequency interference is critical for both accuracy and throughput. For the Five-hundred-meter Aperture Spherical Radio Telescope (FAST), 2D image segmentation pipelines are often too slow at scale, while classical thresholding requires retuning for each site and epoch. We propose PaTRNet, a lightweight parallel Transformer-ResNet architecture operating on 1D spectral energy distribution representations obtained by time-axis accumulation, coupled with channel attention based on Kolmogorov–Arnold networks for dynamic cross-branch fusion. On real FAST data, PaTRNet attains an overall F1 of 0.888 with a per-slice inference latency of 3.04 ms, improving F1 by 14.6% over the strongest traditional baseline (ArPLS-ST) at 2.9 times lower latency, and running 20.6∼48.9 times faster than mainstream 2D segmentation networks while maintaining accuracy. Beyond offline metrics, we apply PaTRNet to FAST observations of the globular cluster NGC 6517 and run a standard PRESTO search on PaTRNet-processed data and raw data. Under identical downstream settings, PaTRNet increases the number of confirmed pulsars from five to six and recovers all five found in the raw-data search, indicating improved candidate visibility and completeness. Furthermore, we tested this pipeline system on the strong pulsar PSR J0837+0610 to ensure that genuine astrophysical signals are not mislabeled.
Li et al. (Tue,) studied this question.