Grain size is a key microstructural parameter influencing material behaviour in a wide range of engineering applications. While microscopy is the standard approach for data acquisition, the traditional methods often face limitations in producing large-size, high-resolution images due to hardware and acquisition constraints. This work presents deep learning-based super resolution techniques as a means to enhance microscopy images beyond their original resolution. Several state-of-the-art architectures are adapted and trained on a dataset comprising various material types. Model performance is benchmarked using quantitative image quality metrics such as structural similarity index (SSIM) [3], Peak Signal-to-Noise Ratio (PSNR) [4], and furthermore visual evaluation for a qualitative perspective. Finally, the top-performing model is further validated through a grain boundary detection task. Here, boundaries detected in the super resolved images are compared against those from the original high-resolution ground truth, evaluation the effectiveness of the super resolution model in reconstructing critical structural features.

References

[1] Y. Chen et al., Optics express, 2024, vol. 32, no. 3, pp. 3316–3328, doi: 10.1364/OE.507017.

[2] L. Song et al., Optics and Lasers in Engineering, 2024, vol. 174, p. 107968, doi: 10.1016/j.optlaseng.2023.107968.

[3] Z. Wang, A. Bovik, H. Sheikh, E. Simoncelli, IEEE Transactions on Image Processing, 2004, vol. 13, no. 4, pp. 600–612.

[4] A. Tanchenko, Journal of Visual Communication and Image Representation, 2014, vol. 25, no. 5, pp. 874–878.