Shape evaluation of highly overlapped powder grains using U-Net-based deep learning segmentation network

With the advancement of electron microscopy, industrial microscale objects are analyzed through image-based characterization. However, the automated and objective assessment of a vast number of images required for quality control is limited by the incomplete segmentation of individual objects in the image. In this study, the scanning electron microscope images of powder grains are selected as target images representing industrial microscale objects. A deep neural network based on the U-Net is developed and trained by manually labeled ground truth. Although the U-Net is a basic network originally devised for biomaterials, the network in this study achieves approximately 90% accuracy and outperforms conventional thresholding methods. However, the boundaries distinguishing individual are not completely classified. The inference results are further processed with morphological operations and watershed algorithms to quantitatively measure grain shapes. Discrepancies in shape parameters between ground truth and network prediction are also discussed.