Automatic microstructure segmentation and quantification with deep learning encoders pre-trained on a large microscopy dataset called MicroNet

A transfer learning approach for the automatic segmentation of microscopy data is presented. Many encoder architectures, including VGG, Inception, ResNet, and others, were trained on ~100,000 microscopy images from 54 material classes to create pre-trained models that learn representations that are more relevant to downstream microscopy analysis tasks than models pre-trained on natural images. The pre-trained encoders were embedded into segmentation architectures including U-Net and DeepLabV3+ to evaluate the performance of models pre-trained on a large microscopy dataset. Each encoder/decoder pair was evaluated on several benchmark datasets. Our testing shows that models pre-trained on a large microscopy dataset generalize better to out-of-distribution data (micrographs taken under different imaging or sample conditions) and are more accurate when training data is limited than models pre-trained on ImageNet. The application of this technique to segment and quantify microscopy features from Ni-superalloys and environmental barrier coatings is demonstrated.