Publications by authors named "Lingxuan Ye"
J Acoust Soc Am
October 2024
Article Synopsis
- This study presents a new method called Learning Direction of Arrival with Optimal Transport (LOT) to improve the estimation of multiple sound sources' directions using sonar signal processing.
- The LOT method treats the direction estimation as a multi-label classification problem and employs an optimal transport loss to better understand angular data, ensuring more accurate predictions and less interference from false targets.
- Additionally, a lightweight data augmentation module is introduced, which enhances the deep learning model's performance by incorporating covariance matrix-related items, and the effectiveness of the LOT approach is validated through experiments on various network architectures and real-world scenarios, particularly with SwellEx-96 data.
J Acoust Soc Am
January 2024
Article Synopsis
- Direction-of-arrival (DoA) estimation is key for sonar signal processing, crucial for tasks like detecting and tracking underwater objects, but estimating orientations of multiple targets accurately remains difficult with current deep learning models.
- To tackle this, the authors improve the permutation invariant training (PIT) approach and introduce two methods: multi-group classification (MC-PIT) and multi-group regression (MR-PIT), allowing a single model to handle training and testing in multi-target situations.
- The study shows that both MC-PIT and MR-PIT outperform traditional multi-label strategies in DoA estimation, and notably, a PIT-trained model demonstrates strong performance on real recorded data even when trained solely on simulated
