摘要
Abstract
To achieve precise detection of defects in manufactured tires,this study proposes a tire defect detection method based on an improved GANomaly network—AAEGAN.First,by replacing the traditional training strategy with a progressive training strategy,the model's convergence speed is significantly improved;second,the encoder is replaced with an adversarial autoencoder(AAE),which,through adversarial training,adds a prior distribution constraint to the latent space;finally,a perceptual loss function is introduced,mark-edly enhancing the model's anomaly detection performance under unsupervised conditions.Comparative and ablation experiments were conducted using a laboratory-made tire dataset.Experimental results show that the AUC increased by 6.3%and the AP increased by 5.1%,validating the method's superior defect recognition capability on a real tire X-ray image dataset.关键词
GANomaly/逐步训练/对抗自编码器/感知损失/异常检测Key words
GANomaly/Progressive training/Adversarial autoencoder/Perceptual loss/anomaly detection分类
信息技术与安全科学
