A Deep Ensemble Classifier for Surface Defect Detection in Aircraft Visual Inspection

Abstract Visual inspection is critical in many maintenance, repair, and overhaul operations and is often the primary defense against premature failure caused by unresolved surface defects. The traditional inspection process is time-consuming and subjective, leading to research into automated systems using computer vision. Several prior methodologies have been developed using convolutional neural networks (CNNs) to classify surface defects; however, these methods often rely on singular models that are sensitive to poor model selection and training errors. Ensembling is a known technique used to minimize the errors of learning algorithms through combining the outputs of multiple models. This paper presents an automated inspection methodology utilizing stacked ensembles of CNNs to classify defects on aircraft surfaces. The proposed framework is evaluated with images obtained from a borescope inspection of aircraft propeller blade bores. It is shown that the ensemble method improves inspection accuracy over conventional single-model deep learning methods. Furthermore, the error reduction provided by the ensemble method reduces false alarms at decision boundaries that minimize missed detections. The proposed method is shown to improve the reliability of automated detection systems, which can avoid catastrophic scenarios on critical systems such as aircraft propellers.