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 7 with the manual annotations and indi- cated that the model approached hu- man-level performance for identifying individual particles. Detecting satellites is a much harder problem, resulting in lower model performance. However, most of the predictions still matched with the annotations, indicating that satellites can consistently be detected in these images. Overlaying the parti- cle and satellite masks to determine the fraction of particles that contain sat- ellites provided a new, objective, and self-consistent method of characteriz- ing the satellite content of powder sam- ples that showed good agreement with the expected trends for images of differ- ent powder samples. CONCLUSIONS The key function of CV is to nu- merically encode the visual information contained in a microstructural image for ML algorithms to find associations and trends. CV/ML systems for micro- structural characterization and analysis span the gamut of image analysis tasks, including image classification, seman- tic segmentation, object detection, and instance segmentation. Applications include: • Visual search, sort, and classifi- cation of micrographs via feature vector similarity. • Extracting information not readily visible to humans, such as chemical composition in SEM micrographs. • Performing semantic segmentation of microstructural constituents with a high accuracy and human-like judgment about what to look for and what to ignore. • Finding all instances of individual objects, even when they impinge and overlap. • Segmenting individual objects to enable new capabilities in micro- structural image analysis. A common characteristic among all of these applications is that they capitalize on the ability of computation- al systems to produce accurate, autono- mous, objective, and repeatable results in an indefatigable and permanently available manner. ~AM&P Acknowledgments This work was supported by the National Science Foundation under grant CMMI-1826218 and the Air Force D3OM2S Center of Excellence under agreement FA8650-19-2-5209. For more information: Elizabeth Holm, professor, Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213, eaholm@andrew.cmu.edu , cmu.edu. References 1. E.A. Holm, et al., Overview: Com- puter Vision and Machine Learning for Microstructural Characterization and Analysis, Materials and Metallurgical Transactions A, 51 (12) , p 5985-5999, 2020. 2. B.L. DeCost, T. Francis, and E.A. Holm, Exploring the Microstructure Manifold: Image Texture Represen- tations Applied to Ultrahigh Carbon Steel Microstructures, Acta Materialia, 133, p 30-40, 2017. 3. B.L. DeCost, T. Francis, and E.A. Holm, High Throughput Quantitative Metallography for Complex Microstruc- tures using Deep Learning: A Case Study in Ultrahigh Carbon Steel, Microscopy and Microanalysis, 25 , p 21-29, 2019. 4. B.L. DeCost, E.A. Holm, Vision-based Methods in Microstructure Analysis, Statistical Methods for Materials Sci- ence: The Data Science of Micro- Fig. 6 — Predicted powder particle (left) and satellite (middle) segmentation masks for SEM images of metal powders used in additive manufacturing. Colors are randomly assigned for visual clarity and do not have physical significance. Sample satellited powder particle (right) detected by overlaying the powder particle and satellite masks. Fig. 5 — Semantic segmentation of microstructural images using a CV/ML system. (a) Segmen- tation of microstructural constituents in an SEMmicrograph of ultrahigh carbon steel. Constitu- ents include network carbide (light blue), ferritic denuded zone (dark blue), Widmanstatten carbide (green), and spheroidite matrix (gold). (b) Segmentation of a tomographic section of an Al-Zn alloy. The solidification dendrite is shown in white on a black background.
Made with FlippingBook
RkJQdWJsaXNoZXIy MjA4MTAy