34 Anomaly Detection with Variational Autoencoders ages are reconstructed with an error. The pixels from which this error originates can indicate the location of an anomaly. In particular, highlighting the difference between an image and its reconstruction can give an indication (to a domain expert observer) where an anomaly may be located. This provides a useful method to speed up the manual checking of potential anomalies in doubtful cases, i.e. examples close to the anomaly decision threshold. It furthermore allows to identify potential improvements to the production process, such as in the visual quality control case, which could prevent future defects. 3.4 Experimental Setup 3.4.1 Datasets MNIST dataset, benchmark setup As a benchmark and test setup, we use the MNIST dataset of handwritten digits (LeCun et al., 1998), which consists of 28×28grey-scale images. The training set consists of 60,000 examples, the test set has 10,000 examples. On this dataset we evaluate ten different binary classification setups, each aiming to classify one digit against the other nine. In other words, in each experiment we consider one of the ten digits to be an anomaly. To evaluate our anomaly framework from Section 3.3, for each anomalous digit, we split up the MNIST dataset as follows: • a training set containing all nine “good” digits from the original MNIST training set (roughly 54,000 examples, depending on the anomalous digit), • a normal test set containing all nine “good” digits from the original MNIST test set (roughly 9,000 examples, depending on the anomalous digit), • an anomaly test set containing all examples of the anomalous digit from the original MNIST training set (roughly 6,000 examples, depending on the digit). This results in ten different anomaly datasets, based on different splits of the original MNIST dataset. Photos of 3D-printed products Our main dataset consists of photos of 3D-printed products. The aim is to detect defects on the outer surface of these products. An example of the full product is shown in Figure 3.2.
RkJQdWJsaXNoZXIy MjY0ODMw