Abstract: Foreign Object Debris (FOD) detection on airport runway is essential for ensuring flight safety since FOD poses a safety threat to aircraft during takeoff or landing. Image Reconstruction based on autoencoder (AE) is a commonly used method for anomaly detection, which assumes that the model trained only using normal samples cannot accurately reconstruct previously unseen anomalies, such as unknown types of FOD. However, deep neural networks have strong generalization capability and accurately reconstruct these anomalies, thereby compromising anomaly detection performance. If the features fed into the decoder do not contain anomalous information, the likelihood of accurately reconstructing anomalies will be significantly reduced. Based on this observation, the Gaussian Mixture Feature Improved Autoencoder (GMFIAE) is proposed for anomaly detection. Given an input image, GMFIAE first obtains the encoding features through the encoder, and then the Gaussian mixture model (GMM) identifies and eliminates potential abnormal features for reconstruction. At the training stage, GMFIAE combines the estimation network to perform end-to-end optimization of AE and GMM to build an accurate normal model. At the test stage, the reconstructed image is obtained from the encoding feature that removes abnormal information, which will tend to be close to the normal image, resulting in a large reconstruction error for anomalies. Furthermore, the multiple hierarchical semantic difference is employed to improve the accuracy of the anomaly detection. The experimental validation is carried out on the runway pavement dataset FODAD from an airport in China and the public anomaly detection dataset MVTec AD. The experimental results demonstrate that the proposed method can achieve accurate anomaly detection.