Recently, deep auto-encoders have been used for the task of anomaly detection in the visual domain. By optimising for the reconstruction error using anomaly-free examples, the common belief is that a corresponding network should fail to accurately reconstruct anomalous regions in the application phase. This goal is typically addressed by controlling the capacity of the network, either by reducing the size of the bottleneck layer or by enforcing sparsity constraints on its activations. However, neither of these techniques does explicitly penalise reconstruction of anomalous signals often resulting in poor detection. We tackle this problem by adapting a self-supervised learning regime that allows the use of discriminative information during training but focuses on the data manifold of normal examples. Precisely, we investigate two different training objectives inspired by the task of neural image inpainting. Our main objective regularises the model to produce locally consistent reconstructions, while replacing irregularities, therefore, acting as a filter that removes anomalous patterns. Our formal analysis shows that under mild conditions the corresponding model resembles a non-linear orthogonal projection of partially corrupted images onto the manifold of uncorrupted (defect-free) examples. This insight makes the reconstruction error a natural choice for defining the anomaly score of a sample according to its distance from a corresponding projection on the data manifold. We emphasise that inference with our approach is very efficient during training and prediction requiring a single forward pass for each input image. Our experiments on the MVTec AD dataset demonstrate high detection and localisation performance. On the texture-subset, in particular, our approach consistently outperforms recent anomaly detection methods by a significant margin.

Within the last decade, neural network based predictors have demonstrated impressive - and at times super-human - capabilities. This performance is often paid for with an intransparent prediction process and thus has sparked numerous contributions in the novel field of explainable artificial intelligence (XAI). In this paper, we focus on a popular and widely used method of XAI, the Layer-wise Relevance Propagation (LRP). Since its initial proposition LRP has evolved as a method, and a best practice for applying the method has tacitly emerged, based on humanly observed evidence. We investigate - and for the first time quantify - the effect of this current best practice on feedforward neural networks in a visual object detection setting. The results verify that the current, layer-dependent approach to LRP applied in recent literature better represents the model's reasoning, and at the same time increases the object localization and class discriminativity of LRP.