Image blur in acquired microscopy images has a negative influence on the quantitative microstructural analysis. Due to the image blur, information of the microstructure is lost, causing analysis methods to provide false measurement results and even domain experts to misinterpret the microstructure. The localization of image blur can help to exclude affected image regions before the analysis and thus saving time and resources. In addition, localization of the image blur provides an overview of the image quality of the acquired sample to the domain expert. In this research, we focus on image blur localization in light-optical microscopy images with a deep learning (DL) approach. To our knowledge, this work is the first deep learning based approach to localize and visualize image blur in materials microscopy images. We considered blur localization as an image segmentation task and conducted this work based on the following three stages: Firstly, we have created a new dataset for blur localization in light-optical microscopy images, as there is no fitting dataset available online at the time of research. Secondly, we trained a U-Net [1] for blur localization. Finally, we conducted a quantitative and qualitative analysis of our DL model and compared it to a frequency based traditional approach provided by Golestaneh et al. [2] and a state-of-the-art DL blur detection model provided by Yang et al. [3].

References

[1] O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015, Bd. vol 9351, pp. 234-241.

[2] S. A. Golestaneh and L. J. Karam, Spatially-Varying Blur Detection Based on Multiscale Fused and Sorted Transform Coefficients of Gradient Magnitudes, 2017, 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017, pp. 596-605, doi: 10.1109/CVPR.2017.71.

[3] X. Xiao, F. Yang, A. Sadovnik, MSDU-Net: A Multi-Scale Dilated U-Net for Blur Detection, 2021, Sensors 2021, 21, 1873. https://doi.org/10.3390/s21051873.