CMU Researchers Propose A Computer Vision-Based Approach With Data-Frugal Deep Learning To Optimize Microstructure Imaging


Materials processing is the process of turning raw materials into final items through a sequence of phases or “unit operations.” The activities entail a series of industrial processes, including various mechanical and chemical methods, which are often carried out in big numbers or batches.

Material processing required extensive analysis and classification of complicated microstructures for quality control. For example, the proportion of lath-type bainite in various high-strength steels affects the material’s characteristics. However, recognizing bainite in microstructural images takes time and money because researchers must first employ two types of microscopy to get a closer look, then rely on their own skills to identify bainitic regions.

Researchers at Carnegie Mellon University have developed computer vision methods for microstructural images that not only require a fraction of the data that deep learning typically requires but also save materials researchers time and money.

Studies suggest that humans find categorization extremely tough; therefore, incorporating a deep learning method will be beneficial.

Their method is extremely similar to that of the larger computer-vision community responsible for facial recognition. To evaluate new photos and classify them, the model is trained on existing material microstructure images. Materials scientists rarely get access to even 10,000 photographs. To that end, the team employed a “data-frugal strategy,” training their model with only 30-50 microscope pictures.

In complex-phase steel, the team experimented with various deep learning algorithms for Laith-bainite segmentation. They were able to attain 90% accuracy, comparable to expert segmentation.

In addition, the team is working on a more cost-effective deep learning method that would just take one image to achieve the same results. Aside from steel, they’re collaborating with a number of research groups to investigate deep learning characterization on a variety of materials.

The researchers hope that by sharing their findings, they will be able to bring this technology to a wider audience in materials science and microstructure characterization.