École Nationale Supérieure de
Techniques Avancées

Résumé

Images perception plays a fundamental role in the tomography-based approaches for microstructure characterization and has a profound impact on all subsequent image processing steps (segmentation and analysis). However, the enhancement of image perception frequently involves the observer-dependence, which translate into user-to-user dispersion and uncertainties in the calculated parameters. This work presents an objective quantitative method, which utilizes convolutional neural networks, for tomographic image quality assessment. Compared to most existing data-driven methods, our method requires less annotations and is more appropriate for tomographic images applications. Different metrics were employed to evaluate the correlation of our predicted scores with the subjective human opinion as well as the segmentation accuracy. The evaluation results from this work demonstrate that our method can be a direct tool that guides the enhancement process and conduct to a reliable segmentation results in respect to the subjective human opinion. As a result, the image processing can turn into a very robust, observer-independent process.