Mr Robert Langtry1
1Research Services Division at Defence Science Technology Group
Detection of pores in materials fabricated using additive manufacturing is key to understand the structural integrity of those materials. Images with infinitesimal pores are highly unbalanced datasets and this complicates the use of machine learning methods.
Application of U-net Neural Networks is ideal for the task of binary segmentation. However, traditional approaches, such as under/oversampling, to unbalanced data rely on input manipulation which cannot be employed for this case. Thus, in order to resolve this issue, I have tested different loss functions from the literature in both synthetic problems and on the aforementioned problem. Using these loss functions the network is able to adapt to imbalanced datasets and gives higher quality predictions of pores than traditional techniques.
Loss functions are able to generate models that suit various statistical measures and allow for the problem of imbalance to be solved without altering the dataset. This neural network solution is agnostic of problem and has use outside of image segmentation tasks as a method for handling unbalanced datasets during training.
In this talk I will go through a variety of loss functions, their links to statistical measures and how that is useful to a data science practitioner in both synthetic examples and the aforementioned pore detection problem.
Robert has a master degree in Computer Science from the University of Melbourne, where he wrote his thesis on Asymmetric numeral system compression. He works as an eResearch Specialist at the Defence Science Technology Group.