The size characterization of particles present in the form of agglomerates in images measured by scanning electron microscopy (SEM) requires a powerful image segmentation tool in order to properly define the boundaries of each particle. In this work, we propose to use an algorithm from the deep statistical learning community, the Mask-RCNN, coupled with transfer learning to overcome the problem of generalization of the commonly used image processing methods such as watershed or active contour. Indeed, the adjustment of the parameters of these algorithms is almost systematically necessary and slows down the automation of the processing chain. The Mask-RCNN is adapted here to the case study and we present results obtained on titanium dioxide samples (non-spherical particles) with a level of performance evaluated by different metrics such as the DICE coefficient, which reaches an average value of 0.95 on the test images.

Instance segmentation model performances is commonly evaluated through challenges using toy datasets such as CIFAR and ImageNet. Although, it is necessary to help improving state-of-the-art models, it remains far from critical industrial application where accurate measurement is required to ensure optimal process control, quality assurance and safety. This paper compares the performance of ready-to-deploy state-of-the-art instance segmentation models, specifically region and query-based models, across COCO and task-specific designed metrics in the context of the estimation of aggregated TiO₂ particle size distribution from scanning electron microscopy (SEM) measurements. The numerical comparison is conducted on custom datasets, addressing well-known challenges of SEM microscopists: high resolution images showing a massive number of overlapping instances, images measured with diverse imaging parameters (operator control) or the presence of a mixture of aggregated particles. Our findings highlight the strengths and limitations of each model category, providing insights into their suitability for high-stakes industrial deployment. Finally, we propose a customized inference procedure that leverages the performance of transformers on images that contain a small number of instances within dense and clustered scenes.