BSE SEM denoiser
U-Net model for denoising back-scattered SEM images.
Scanning Electron Microscope (SEM) Backscatter Electron (BSE) imaging is often used to suppress charging and topography in images. Many BSE images can show a significant amount of noise.
The BSE denoiser model presented here consists of a convolutional neural network (CNN) with U-Net architecture. This network was trained on a wide range of BSE images with artificially induced Poisson noise.
The model was tested on a broad variety of BSE-SEM images such as metals, neurological tissue, reinforced polymers, and batteries. The output data was on par with and often outperformed classical denoising filters such as Non-Local Means (NLM) and Block Matching 3D filtering (BM3D).
- model_sem_bse.json = model architecture,
- model_sem_bse.hdf5 = trained weights,
- model_sem_bse.py = pre-processing functions required to connect the model to Amira-Avizo/Pergeos grayscale images.
- a test image taken with different dwell times
The model can be applied to any grayscale image or volume, using the Deep Learning Prediction module.
Note that, as for any noise reduction filters, the performance is related to the type of noise actually affecting the images. Also, as with any Deep Learning model, artefacts may be introduced if the data is too different from the training database. The results always should be considered with caution.
Download the Xtra compatible with versions 2021.1 and 2021.2 HERE.