Foreign Object Debris Detection for Airport Pavement Images Based on Image Random Masking and Semantic Feature Matching

Foreign object debris (FOD) detection on airport runways plays a crucial role in ensuring aircraft flight safety as it severely threatens the safety of aircraft takeoff or landing. Interpreting FOD as anomaly on airport runways, the unsupervised anomaly detection methods based on image reconstruction can detect various types of FOD. However, the existing image reconstruction methods are still imperfect due to the excellent generalization of deep neural networks, the presence of low-contrast anomalies, and the instability of reconstruction error. This article proposes a structural grid masking anomaly detection (SGMAD) network for unsupervised anomaly detection, which removes partial regions in the image via multiscale structural grid masks (MSGM) and restores the complete original information. The drawbacks of image reconstruction methods can thus be addressed. Besides, feature-level reconstruction error is introduced to enlarge the error gap between normal and abnormal data, thereby improving the accuracy of low-contrast anomaly detection. To achieve this, semantic feature matching (SFM) is proposed, introducing an adversarially trained discriminator to explore semantic feature representation. The experiments conducted on the airport runway dataset FODAD and the public anomaly detection dataset MVTec AD demonstrate that SGMAD outperforms state-of-the-art methods and has excellent potential for real industrial applications.