Reconstruction-based methods have recently shown great promise for anomaly detection. We here propose a new transform-based framework for anomaly detection. A selected set of transformations based on human priors is used to erase certain targeted information from input data. An inverse-transform autoencoder is trained with the normal data only to embed corresponding erased information during the restoration of the original data. The normal and anomalous data are thus expected to be differentiable based on restoration errors. Extensive experiments have demonstrated that the proposed method significantly outperforms several state-of-the-arts on multiple benchmark datasets, especially on ImageNet, increasing the AUROC of the top-performing baseline by 10.1%. We also evaluate our method on a real-world anomaly detection dataset MVTec AD and a video anomaly detection dataset ShanghaiTech to validate the effectiveness of the method in real-world environments.