Foreign Object Debris Detection Based on Gaussian Mixture Autoencoder of Pre-trained Features

Abstract

In this study, a novel anomaly localization method called Gaussian Mixture Autoencoder of Pre-trained Features (GMAPF) is proposed to perform foreign object debris (FOD) detection in the field of aviation. GMAPF utilizes the pre-trained deep convolutional neural network to establish multi-hierarchical feature representations, which are then fed into the deep autoencoder for dimensionality reduction and learning of low-dimensional embedding for each pixel of an image. The distribution of the normal pixel embedding is then modeled by Gaussian mixture model (GMM). Besides, instead of Expectation-Maximization (EM), GMAPF leverages a multi-layer perceptron to learn the parameters of GMM. Therefore, GMAPF could simultaneously optimize the parameters of the deep autoencoder and GMM in an end-to-end way. Many experiments are done on a newly collected dataset FOD, and the experimental results demonstrate the validity of GMAPF.