Merged-Sampling Mask R-CNN With Random Proposal Expansion for Particle Measurement of SEM Images of Molecular Sieve Catalysts

Abstract

Scanning electron microscope (SEM) images of molecular sieve catalysts contain information about the shapes and sizes of these particles. SEM image measurement is a crucial step in the evaluation of the catalytic performances. Instance segmentation methods, such as Mask R-CNN are effective in automatically analyzing SEM images. However, their performance is limited in small datasets. Although overfitting caused by small data sets can be addressed through data augmentation at the image level, the application of Mask R-CNN still needs further improvement for generalization enhancement. In this paper, two techniques for Mask R-CNN are proposed to alleviate overfitting during the training on small datasets. First, merged sampling on the region proposal network simultaneously samples proposals with high and low scores in order that the head networks can be exposed to more diverse proposal samples. Second, random proposal expansion enhances the diversity of samples provided to the mask-branch of Mask R-CNN. These two techniques can be viewed as data augmentation at the instance level. Experiments on a small SEM image dataset showed that merged-sampling Mask R-CNN with random proposal expansion improved about the average precision (AP) by 5%, compared with the original Mask R-CNN. Overfitting on the small dataset is effectively controlled using the proposed methods. Hence, the results of the particle measurements of the industrial SEM images improved.