A process created by Codecentric generates anomalous training data to detect anomalies in the quality control of surfaces using AI. One or more so-called obfuscators add artifacts or effects to error-free images that interrupt and disrupt the structure of the surface. A U-Net autoencoder and perceptual loss are used to try to remove these synthetic errors from the images. The anomaly becomes visible based on the difference between the image with the anomalies and the result of the U-Net autoencoder, since both images only differ in the locations of the anomalies. The dataset used contains images of 15 different product surfaces, four of which are examined in more detail. The dataset contains error-free and error-prone images. There are also masks that show the position and outline of the errors and are used to evaluate the process.
