A comparison of deep-learning-based inpainting techniques for experimental X-ray scattering

Area of Science:

• Materials Science
• Data Science
• Image Processing

Background:

• Experimental X-ray scattering data often contains gaps due to experimental limitations.
• Accurate reconstruction of missing data is crucial for reliable scientific analysis.

Purpose of the Study:

• To implement and evaluate deep learning-based image inpainting techniques for reconstructing gaps in X-ray scattering data.
• To compare the performance of novel deep learning architectures against traditional inpainting algorithms.

Main Methods:

• Utilized deep learning neural network architectures: convolutional autoencoders, tunable U-Nets, partial convolution neural networks, and mixed-scale dense networks.
• Reconstructed missing pixel intensities in experimental scattering images.
• Evaluated reconstruction accuracy using mean absolute error and correlation coefficient metrics against ground-truth data.

Main Results:

• Deep learning methods significantly outperformed traditional algorithms like biharmonic functions.
• Tunable U-Net and mixed-scale dense network architectures demonstrated the highest reconstruction performance.
• Achieved correlation coefficient scores exceeding 0.9980 for the best-performing models.

Conclusions:

• Deep learning image inpainting is a powerful tool for handling missing data in X-ray scattering experiments.
• Advanced neural network architectures offer superior accuracy for data reconstruction.
• The proposed methods enhance the reliability and completeness of experimental scattering datasets.