Abstract: In response to the difficulties in collecting defect samples in real- world industrial production scenarios, this paper proposes a defect detection model based on improved generative adversarial networks and deep feature repair. This paper designed a defect fitting module and incorporated a prediction loss function to improve detection accuracy. The pre- trained model and the feature fusion module were used to fully mine the depth features of input images. This paper designed and modified the skip connection structure of generators in GANs, while incorporating self attention mechanism into the discriminator to improve the model's ability to repair deep features. Experimental results on the MVTec AD dataset show that the average AUC reached 0.964. Compared with the sub- optimal model, it was increased by 2.3 percentage points. Experimental results on the self- made casting industry dataset show that the average AUC reached 0.988 and the average detection speed was 260ms per piece. It verifies that the proposed model is suitable for real- world industrial production scenarios.