Comparison of deep learning methods for image deblurring on light optical materials microscopy data

The robustness of image processing and machine learning algorithms for object and anomaly detection, image segmentation or failure analysis tasks is strongly influenced by the image quality. Image blur is still a problem in microscopy because high quality images of non-planar samples with high resolution are not feasible even with high technical and manual effort. In this paper we evaluate deep neural network models for image deblurring on light optical microscopy data. We present a new image deblurring dataset with sharp ground truth images and a variation of different out-of-focus blur and vibration blur images. We show that image quality enhancement using deep learning methods has great potential in microscopy-based failure analysis. The best method achieved an improvement for the PSNR metric from 31.24 to 35.19, for the SSIM metric from 0.7981 to 0.9472 and for the IoU score from 0.845 to 0.944 on the given test dataset.