Computer Vision News - November 2021

51 Anomaly detection in medical imaging with ... Examples of dataset with abnormality vs normality comparisons. CIFAR10 dataset images, with SVHN digit dataset in the row below, followed by histopathological and chest X-ray images from the NIH dataset. DEEP PERCEPTUAL AUTOENCODER Autoencoder-based approaches rely on the fact that they can learn shared patterns of the normal images and then restore them correctly. The key idea of the proposed method is to simplify the learning of these common factors inherent to the data, by providing a loss function that measures "content dissimilarity" between the input and the output. The progressive training idea is explored further to improve the expressive power of the autoencoder. The reasoning is that the pipeline gradually grows the "level" of the "perceptual" information in the loss function. The paradigm has many hyperparameters, therefore tuning is essential to ensure detection quality. Labels can be used during the model setup to create a weakly-supervised training paradigm, where a low number of labelled anomalous examples of a limited variation is available. has been extensively studied. Domains including fraud detection, cyber-intrusion detection, as well anomalies in videos and financial analytics are some of the latest fields where research is active. Distribution based methods try to predict if the new example lies in the high-probability area or not.