To ease visual inspections of exterior aircraft fuselage, new technical approaches have been recently deployed. Automated unmanned aerial vehicles (UAVs) are now acquiring high-quality images of aircraft in order to perform offline analysis. At first, some acquisitions are annotated by human operators in order to provide a large dataset required to train machine learning methods, especially for critical defect detection. An intrinsic problem of this dataset is its extreme imbalance (i.e., there is an unequal distribution between classes). The rarest and most valuable samples represent few elements among thousands of annotated objects. Deep learning (DL)-only based approaches have proven to be very effective when a sufficient amount of data are available for each desired class, whereas few-shot learning (FSL)-dedicated methods (matching network, prototypical network, etc.) can learn from only few samples. In a previous work, those approaches were compared on our applicative case and it was demonstrated that combining DL model and prototypical neural network in a hybrid architecture improves the results. We extend this work by questioning the interface between models in such a hybrid architecture. We show that by carefully selecting the data from the well-represented class when using FSL techniques, it is possible to enhance the previously proposed solution.
In order to ease visual inspections of exterior aircraft fuselage, new technical approaches have been recently deployed. Automated UAVs are now acquiring high quality images of the aircraft in order to perform offline analysis. At first, some acquisitions are annotated by human operators in order to provide a large dataset required to train machine learning methods, especially for critical defects detection. An intrinsic problem of this dataset is its extreme imbalance (i.e there is an unequal distribution between classes): the rarest and most valuable samples represent few elements among thousands of annotated objects. Deep Learning-only based approaches have proven to be very effective when a sufficient amount of data is available for each desired class, whereas less complex systems such as Support Vector Machine theoretically need less data, and few-shot learning dedicated methods (Matching Network, Prototypical Network, etc.) can learn from only few examples. Those approaches are compared on our applicative case. Preliminary results show the existence of empirical frontiers in term of training dataset volume that indicate which approach might be favored. Based on those results, we propose a method to combine different approaches in order to achieve best performances on categorical accuracy, with special attention paid to underrepresented category.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.