Feb/March_AMP_Digital

A D V A N C E D M A T E R I A L S & P R O C E S S E S | F E B R U A R Y / M A R C H 2 0 2 1 1 6 annotations, and certainly adequate for quantitative analysis. Besides the excel- lent performance, the CV/ML system is also fast, autonomous, objective, and repeatable, enabling the high through- put necessary for applications such as 3D reconstruction or quality control. An additional benefit of this ap- proach is the ability to capture human- like judgments about image features. For instance, in Fig. 5a, the spheroid- ite matrix constituent, comprised of spheroidite particles in a ferrite matrix, is segmented as a single constituent (or- ange). Likewise, in Fig. 5b, the system learns to ignore sample preparation ar- tifacts such as the sample edge, pores, and the circular beam spot at the cen- ter of the image. Conventional segmen- tation methods would be challenged to handle these complex features. It is this capacity for learning what to look for and what to ignore that distinguish- es the CV/ML approach to semantic segmentation. OBJECT DETECTION AND INSTANCE SEGMENTATION Object detection entails locat- ing each unique object of its kind in an image, i.e., finding each individual precipitate in a micrograph. Instance segmentation extends this technique to also generate segmentation masks for each individual object. Specialized CNNs have been developed for object detection and instance segmentation [22] . As in the case of semantic segmenta- tion, transfer learning allows models trained on natural images to be adapt- ed to materials science applications. For example, the presence of small satellite particles is known to affect the flowability of metal powders used in ad- ditivemanufacturing. ACV/ML approach utilizing object detection and instance segmentation demonstrated the abili- ty to identify individual powder parti- cles and their satellites in dense powder images. Tedious manual annotation yielded five and ten labeled images for powder particles and satellites, respec- tively. The CV/ML systemwas trained on these images, and sample predictions are shown in Fig. 6. The powder parti- clemasks showed very good agreement images via transfer learning. For exam- ple, the PixelNet CNN [20] trained on the ImageNet database of natural images [14] has been used to classify pixels accord- ing to their microstructural constituent as shown in Fig. 5. In Fig. 5a, the system was trained using 20 hand-annotated Fig. 4 — Classification results for steel inclusion composition from a database of 2543 backscattered SEM image patches (example images are shown to the right). The prediction accuracy for each inclusion type is shown along the diagonal; overall accuracy is about 76%. Fig. 3 — Measurement of average grain size from a database of 15,213 synthetic polycrystalline microstructures using deep regression. The red line corresponds to perfect accuracy. Inset shows an example microstructure. images from the UHCS micrograph da- tabase [3] , and in Fig. 5b, the system was trained on 30 hand-annotated images from a set of tomographic slices of an Al-Zn solidification dendrite [21] . In both cases, the predicted segmentations are arguably equal in quality to the human