Abstract: In response to the challenges of acquiring and labeling defect data on conveyor belts, as well as the low accuracy of deep learning-based conveyor belt defect detection methods due to unstable factors and data fluctuations in working environments, this study proposed a surface defect detection model based on adversarial repair networks. The model primarily consisted of a generator with an autoencoder structure and a Markov discriminator. During the training phase, simulated surface defect images of the conveyor belt were input into the generator to obtain reconstructed images without simulated defects, enhancing the model's ability to generalize to unknown defects. The original undamaged conveyor belt images, reconstructed images, and simulated surface defect images were input into the Markov discriminator, and feature maps were obtained through a residual network, improving the model's detection capability for subtle defects. In the detection phase, the test image was input into the trained generator to obtain the reconstructed image, and the trained Markov discriminator was used to extract feature maps from both the test image and the reconstructed image. The anomaly score was calculated based on the mean squared error between the feature maps of the test image and the reconstructed image, as well as the maximum value of the feature map of the test image, and compared with a set threshold to determine whether the test image contained defects. Experimental results showed that the area under the receiver operating characteristic curve (ROC-AUC) of this model reached 0.999, the area under the precision-recall curve (PR-AUC) reached 0.997, and the detection time for a single image was 13.51 ms, which could accurately locate the positions of different types of defects.