CNN-Transformer with Sigmoid Label for High-Resolution Sub-Degree DOA Estimation
A CNN-Transformer with sigmoid labels reaches sub-degree DOA accuracy on a 0.25-degree grid.
The paper proposes a sigmoid label for deep-learning direction-of-arrival estimation to reduce grid quantization error. A CNN-Transformer trained with summed angle-shifted sigmoids, then differentiated, produces sharp spectral peaks. With a 0.25-degree grid it reports sub-degree accuracy and 1-degree resolution, outperforming Gaussian, hard 0-1, and binary cross-entropy labels under the same backbone.
- Sigmoid labels encode source angles as shifted monotonic functions.
- Differentiating CNN-Transformer outputs yields sharp DOA peaks.
- A 0.25-degree grid yields sub-degree accuracy and 1-degree resolution.
- Sigmoid labels beat Gaussian, hard 0-1, and BCE labels.
From the original article
1 min read at arxiv.orgThis paper addresses the grid quantization error of deep-learning (DL)-based direction-of-arrival (DOA) estimation. Classification- based DOA methods discretize the continuous angular space into a grid, so that with a peak-based read-out, the accuracy and resolution are limited by the grid step and by the width of the label peak.
By Lifeng Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.