Elsevier

Computational Materials Science

Volume 126, January 2017, Pages 438-445
Computational Materials Science

Characterizing powder materials using keypoint-based computer vision methods

https://doi.org/10.1016/j.commatsci.2016.08.038Get rights and content

Highlights

  • A computer vision model classifies a large data set of powder images.

  • Visually similar, quantitatively distinct images are sorted with 90% accuracy.

  • The computer vision accuracy is comparable to traditional measurement methods.

  • This work is applicable to feedstock powders for additive manufacturing processes.

Abstract

We applied the bag of visual words model for visual texture to a dataset of realistic powder micrograph images drawn from eight closely related particle size distributions. We found that image texture based powder classification performance saturates at 89±3% with 640 training images (80 images per class). This classification accuracy is comparable to classification using conventional segmentation-based particle size analysis. Furthermore, we found that particle size distributions obtained via watershed segmentation are generally not statistically equivalent to the ground truth particle size distributions, as quantified by the two-sample Kolmogorov-Smirnov test for distribution equivalence. We expect image texture classification methods to outperform particle size analysis for more challenging real-world powder classification tasks by capturing additional information about particle morphology and surface textures, which add complexity to the image segmentation task inherent in particle size distribution estimation.

Introduction

The discipline of materials science has long history of quantifying microstructural images to interpret structure–properties–processing relationships. However, many of the microstructures found in technologically important materials systems are complex and do not map well onto a reductionist segment-and-measure paradigm, at least under the constraints of our current understanding of mesoscale physics. Where explicit mesoscale materials models are not available, there has been interest in applying more flexible image representations to microstructure data in an attempt to circumvent these limitations [1], [2], [3], [4], [5]. For example, we recently demonstrated that the ‘bag of visual words’ (BOVW) image texture representation can be used to automatically differentiate between qualitatively different microstructures [5].

As materials scientists we are ultimately interested in learning quantitative structure properties and processing structure mappings. We hope that image texture characterization methods can aid us in this goal. A critical outstanding question in this approach is the problem of physical size: can scale-invariant texture features (like the BOVW approach) support quantitative structure properties mappings that depend strongly on relative feature sizes? In this work, we move towards quantitative microstructure characterization by focusing on microstructures that vary primarily in their physical dimensions. Here we focus on a simple model system: powder materials with different particle size distributions (PSD).

The size distribution and morphology of powder materials play important roles in powderbed fusion additive manufacturing (AM) [6], [7], [8] and other powder metallurgy applications [9]. This class of AM techniques enables rapid building of parts with complex geometries in a layer-by-layer process where a powder feedstock material is locally fused using a computer-routed laser or electron beam [10], [11]. The PSD and morphology can play a large role in powder rheology [6], [8], and can affect the density distribution in the powder bed, which in turn can alter energy absorption and thermal conductivity [12], [13]. Together these factors influence both the final microstructure of the built part, and its properties such as porosity, hardness, mechanical strength, and surface roughness [14], [15].

Understanding how powder character affects the build process and final properties is especially important when recycling the unused portion of the powder bed, because the powder character has been shown to change with recycling [14], [15], [16]. Unused powder can contain distorted partially melted particles, sintered particle clusters, or solidified melt-pool ejecta [14], [15]; recycled powders are typically sieved to remove these, and to ensure that the recycled powder has a maximum particle size consistent with the powder bed thickness. Recent powder rheometry studies indicate that sieving may not always mitigate the adverse effects of recycling; powder flowability may decrease [15], [17] or increase [14], [16] after recycling. For these reasons, it is important to characterize the morphology of AM powder feedstock material, and to couple this with an understanding of how powder morphology influences the build process [18], [15], [17].

A wide array of characterization techniques has been applied to AM powder feedstock materials [18], including laser PSD measurements [19], X-ray computed tomography (XCT) to measure three dimensional particle sizes and shapes, X-ray diffraction for phase identification, scanning electron microscopy to qualitatively study surface structure, and energy-dispersive spectroscopy (EDS) and X-ray photospectroscopy (XPS) to measure chemical composition. Importantly, the PSD alone does not provide information about powder morphology that may be important for understanding the behavior of the powder [15]. Consider the gas-atomized powders shown in Fig. 1. Some of the powder particles are non-spherical and appear to have non-trivial surface roughness. Furthermore, many of the particles obscure and occlude each other, and smaller particles tend to agglomerate. These morphological features may not be adequately resolved by laser particle size distribution measurements [18] or region-based image analysis techniques [17], and it may be impractical and/or cost-prohibitive to conduct XCT and powder rheometry experiments regularly in a production setting. Therefore it is of interest to develop powder characterization techniques based on image texture analysis to complement the suite of analytical tools currently in use.

In this study, we investigate the capability of image texture methods, specifically the BOVW model, to capture quantitative differences in microstructure, where the primary aspect of interest is the relative size of microstructure features. Using 3D rendering to create realistic powder micrographs, we generated a dataset of 2048 synthetic micrographs representing eight different powder PSDs. This dataset provides a realistic materials science application, which we use to exercise and evaluate the BOVW representation for microstructure characterization and to delineate the amount of microstructure information required to support a data-driven powder characterization approach. We also evaluate particle size measurements obtained through a conventional segmentation-based method by comparing to the ground truth particle sizes in the rendered images. We conclude that image texture methods, together with rheological and build-performance studies, offer a promising approach to evaluate and qualify powder feedstock for AM processes.

Section snippets

Synthetic powder micrographs

We created a dataset of realistic powder micrographs; Fig. 2 shows example renderings. This microstructure synthesis task was performed using Blender [20], an open source computer graphics suite used for 3D modeling, rendering, animation, and scientific visualization; the scripts and texture resources used to generate this dataset are included in a data-in-brief summary [21]. In the present study, we limit ourselves to powders consisting of spherical particles with spatially uncorrelated

Bag of visual words

To quantitatively evaluate how well the BOVW representation can characterize powder micrographs, we performed χ2-kernel SVM classification [39], [40]. We use half of our dataset (128 images from each of the eight generating distributions) to tune model parameters (the number of visual words in the dictionary k and the SVM regularization parameter C) via 5× 8-fold cross-validation, randomly partitioning the training set into 7 folds for training and one fold for testing, repeated five times. We

Discussion

For this synthetic powder micrograph dataset, the BOVW method yields comparable classification results to segmenting and measuring sphere-equivalent particle size distributions. This demonstrates that the BOVW model is not necessarily limited to applications involving qualitatively different microstructures, but has the potential to provide quantitative microstructural insight. We believe that the BOVW features for this particular synthetic powder dataset effectively capture the shape of the

Conclusions

We apply the bag of visual words (BOVW) computer vision technique to obtain a rich microstructure representation that can be used to explore process–structure–properties mappings. We find that BOVW classification performance saturates at 89±3% with 80 synthetic powder micrograph instances for each of eight generating particle size distributions. The BOVW classification accuracy is comparable to conventional segmentation-based particle size analysis. Though watershed segmentation yields

Acknowledgements

We gratefully acknowledge funding for this work through the National Science Foundation Grant Nos. DMR-1307138 and DMR-1507830, and through the John and Claire Bertucci Foundation. A.D. Rollett, R. Cunningham, and H. Jain graciously supplied the experimental powder micrographs in Fig. 1. Thanks to VLFeat [36], Scikit-Learn [41], Blender [20], and ImageJ [42] for the awesome open source code. We are also grateful for helpful and inspirational discussions with Prof. Abhinav Gupta and Xinlei Chen

References (48)

  • A. Simchi

    The role of particle size on the laser sintering of iron powder

    Metall. Mater. Trans. B

    (2004)
  • John A. Slotwinski et al.

    Metrology needs for metal additive manufacturing powders

    JOM

    (2015)
  • Randall M. German

    Powder Metallurgy and Particulate Materials Processing: The Processes, Materials, Products, Properties, and Applications

    (2005)
  • D.D. Gu et al.

    Laser additive manufacturing of metallic components: materials, processes and mechanisms

    Int. Mater. Rev.

    (2012)
  • William E. Frazier

    Metal additive manufacturing: a review

    J. Mater. Eng. Perform.

    (2014)
  • Mahesh Mani et al.

    Measurement Science Needs for Real-time Control of Additive Manufacturing Powder Bed Fusion Processes

    (2015)
  • Jamie Clayton et al.

    The application of powder rheology in additive manufacturing

    JOM

    (2015)
  • H.P. Tang et al.

    Effect of powder reuse times on additive manufacturing of ti-6al-4v by selective electron beam melting

    JOM

    (2015)
  • A. Strondl et al.

    Characterization and control of powder properties for additive manufacturing

    JOM

    (2015)
  • J.A. Slotwinski et al.

    Characterization of metal powders used for additive manufacturing

    J. Res. Natl. Inst. Stand. Technol.

    (2014)
  • Vincent A. Hackley et al.

    Particle Size Analysis by Laser Diffraction Spectrometry: Application to Cementitious Powders

    (2004)
  • Roland Hess

    The Essential Blender: Guide to 3D Creation with the Open Source Suite Blender

    (2007)
  • B.L. DeCost, E.A. Holm, A large dataset of synthetic SEM images of powder materials and their ground truth 3D...
  • Svetlana Lazebnik et al.

    A sparse texture representation using local affine regions

    IEEE Trans. Pattern Anal. Machine Intell.

    (2005)
  • Cited by (0)

    View full text