R-DCNN: Low-Complexity Deep Learning for Periodic Signal Denoising
A recent study introduces R-DCNN, an efficient deep learning technique designed for denoising periodic signals and waveform estimation. This method integrates dilated convolutional neural networks with a resampling process, allowing it to function effectively within limited power and resource parameters. It is specifically aimed at signals that exhibit varying fundamental frequencies and requires just one observation for training. By employing lightweight resampling to synchronize time scales, the same network weights can be utilized across multiple frequencies. Despite its low complexity, R-DCNN delivers performance on par with leading methods. This research is available on arXiv (2604.21651) and is relevant to fields such as speech, music, medical diagnostics, radio, and sonar.
Key facts
- R-DCNN combines dilated CNNs with resampling for periodic signal processing.
- Designed for low computational resources and strict power constraints.
- Requires only a single observation for training.
- Resampling aligns time scales to reuse network weights across frequencies.
- Achieves performance comparable to state-of-the-art methods.
- Published on arXiv with ID 2604.21651.
- Applicable to speech, music, medical diagnostics, radio, and sonar.
- Proposes a method for denoising and waveform estimation.
Entities
Institutions
- arXiv