Pulsar-vision-language Model: A Visual Reasoning Model for Pulsar Candidate Identification

Le résumé fourni par la source

Abstract The rapid advancement of radio telescopes has resulted in a massive increase in pulsar candidate data, rendering manual inspection infeasible. Existing convolutional neural network–based identification models often struggle with variations in diagnostic plots across different telescopes. We propose a pulsar-vision-language model (pulsar-VLM) for pulsar candidate identification, which is built upon pretrained multimodal large language models. By integrating visual features from diagnostic subplots with semantic textual descriptions, the model performs geometric morphological reasoning in a manner analogous to expert astronomers. To enhance the diversity of the training data, we augment the pulsar candidate data by selecting diagnostic plots from radio-frequency interference (RFI) signals that visually resemble pulsar candidates. Additionally, we construct an evaluation benchmark consisting of 13,024 real pulsar candidate diagnostic plots and 171,493 RFI samples from five commonly used radio telescopes, covering both standard and nonstandard diagnostic plots. Experimental results on the benchmarks demonstrate that pulsar-VLM achieves a 94.2%–100% recall across different radio telescopes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.