Low-resolution and noisy one-dimensional signals including Gaussian peaks (e.g., molecular spectra, LiDAR histogram data, and other data) are denoised and up-sampled using a deep neural network model. The model is similar to the one proposed in [1] and based on a stack of residual blocks [2] and a point-shuffle method inspired by pixel shuffle [3].
In the example below, signals including 128 data points (blue curve) are up-sampled to 512 (red curve) and denoised at the same time.
Below are a few examples of zoomed-in versions to show up-sampling and denoising effectrs better:
In the examples shown above, the number of residual blocks, which can be easily changed in the model.py script, is set to be only 2 for quick demo purposes. In this case, the number of trainable parameters is 249,476.
generate_data.py: Synthetic dataset can be created by running this script. Low-resolution and noisy data and high-resolution and noise-less data are generated. The number of Gaussian peaks, the number of data points for high- and low-resolution data, peak widths, noise levels, and amplitudes can be set.model.py: A model is defined.loss.py: A custom loss function is defined.main.py: This is a script for training and evaluating the model. You can change CONFIG to do some trainings, evaluations, and experiments.
See [1] for further details. Please note that I am not the author of the paper. I made this repository for learning purposes.
[1] Gangping Liu and Jun Ke, Full-waveform LiDAR echo decomposition based on dense and residual neural networks, Applied Optics, vol. 61, 9, pp. F15-F24, 2022
[2] Paper With Code, "Residual Block," paperswithcode.com. [Online]. Available: https://paperswithcode.com/method/residual-block [Accessed: 31 May, 2024]
[3] Paper With Code, "Pixel Shuffle," paperswithcode.com. [Online]. Available: https://paperswithcode.com/method/pixelshuffle [Accessed: 31 May, 2024]