Abstract: In the field of anomaly detection, Generative Adversarial Nets (GAN) and Autoencoder (AE) have achieved better application results in recent years. However, the existing GAN-based anomaly detection models generally suffer from poor reconfiguration ability. To address this issue, this paper proposes a two-discriminator GANomaly network model, in which the global discriminator is used to improve the reconstruction ability of images and the local discriminator is used to improve the encoding ability at the spatial level. The proposed method is validated on the MvTec dataset and the homemade tire X-ray image dataset, respectively. The experimental results show that the proposed method can effectively improve the reconstruction ability of the model, reduce the anomaly score threshold, and improve the accuracy of anomaly detection.