Weak celestial source fringes detection based on multi-task learning and pseudo soft threshold residual denoising network

Detecting low signal-to-noise ratios weak celestial source signals from large volumes of astronomical radio data is a significant and challenging task. Current mainstream approaches predominantly rely on manual processing, resulting in low efficiency. While a few deep learning-based methods have been applied, they typically utilize generic techniques, leading to suboptimal detection accuracy. To address these challenges, this paper proposes a novel method for detecting weak celestial source signals, tailored to the unique characteristics of celestial source data. The method integrates multi-task learning with pseudo-soft threshold residual denoising. Firstly, transfer learning is introduced to leverage a pre-trained model for extracting features from celestial source fringes and performing signal recognition. Multi-task learning is employed to enhance detection efficiency and reduce the false detection rate. Secondly, a novel pseudo-soft threshold function is proposed, and a corresponding pseudo-soft threshold residual denoising network is developed to automatically learn the optimal threshold and eliminate noise features. Additionally, a multi-layer fusion feature pyramid network is proposed to improve the extraction of features from weak celestial source fringes. Simulated data, generated based on the parameters of the Tianlai radio telescope observation system, is used to construct a training dataset. The performance of the proposed algorithm is evaluated using both simulated and real observational data. Experimental results demonstrate that the proposed method achieves satisfactory recognition accuracy, providing significant benefits for astronomers in detecting weak celestial source signals from extensive radio observation data. The code of this work will be available at https://github.com/YanRuiqing/MTL-PSTRD to facilitate reproducibility.