"Deep learning for denoising SEM Images" by Khaled Al Saih

Seminar by Khaled Al Saih

Abstract

Microscopy image investigation mostly requires the segmentation of targeted elements, but training data is typically tight and challenging to obtain due to the unavailability of ground truth images. Here we trained a deep learning end-to-end model named DenoiSeg on synthetic data generated using physically-based noise models. Denoiseg is an extension of Noise2Void. Unlike other architectures, DenoiSeg takes into account the segmentation and denoising information which are trained jointly in an encoder-decoder style. We generated three synthetic datasets which mimicked the SEM real images, called 1kV-10pA, 2kV-10pA, and 2kV-38pA. Each dataset contains 3000 images and during the training, the augmentation is performed. Data trained is validated using 4k fold cross-validation. The performance evaluation of the three trained networks is performed using two metrics, namely, Peak signal-to-noise rate (PSNR) and Similarity index measurement (SSIM). We obtained first results: the improved PSNR for the three datasets are 13.97, 13.68, and 13.78, respectively ; the Improved SSIM for the datasets are 0.48, 0.46, 0.5. To sum up, these results are promising using the synthetic data and it still can be improved.

This seminar will be held in english.

Lien Cisco Webex :   https://ujmstetienne.webex.com/meet/ImageSCV