Abstract

The fourth industrial revolution has brought many benefits to the mechanical engineering world. Through data revolution, defect segmentation in complex XCT images can now be automated in a shorter time, while achieving more accurate results. Prior work[1] proves that a deep learning approach can extract pore from every voxel of a 3D XCT data, and calculate its porosity with a high accuracy. However, it was established that training a deep learning model with limited data can cause the model to overfit, or have inferior segmentation performance than models that are trained with a larger dataset. This is a common issue in all sorts of data-driven inspection. Oftentimes, obtaining raw data and annotating these raw data for machine learning inspection can take significant resources, shifting the focus away from the network design and training. We present XCTPore, an open source database of X-ray CT images, containing 2D slices and 3D volumetric data of X-ray CT scanned additively manufactured components, with varying porosity and image quality. This database is currently maintained by the Advanced Remanufacturing and Technology Centre(ARTC) and is open to contributions from industry practitioners, and academics. The content of the database can be queried through our python code found in our GitHub link. This database also comes with an open-source implementation of unsupervised porosity labelling function in a bid to automatically label pores without the need for human intervention. This unsupervised labelling is done using STEGO, to aid developers to label the images, to explore novel approaches.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call