Anomaly Detection with Deep Perceptual Autoencoders

  • 2020-10-15 17:26:07
  • Nina Tuluptceva, Bart Bakker, Irina Fedulova, Heinrich Schulz, Dmitry V. Dylov
  • 0


Anomaly detection is the problem of recognizing abnormal inputs based on theseen examples of normal data. Despite recent advances of deep learning inrecognizing image anomalies, these methods still prove incapable of handlingcomplex medical images, such as barely visible abnormalities in chest X-raysand metastases in lymph nodes. To address this problem, we introduce a newpowerful method of image anomaly detection. It relies on the classicalautoencoder approach with a re-designed training pipeline to handlehigh-resolution, complex images and a robust way of computing an imageabnormality score. We revisit the very problem statement of fully unsupervisedanomaly detection, where no abnormal examples at all are provided during themodel setup. We propose to relax this unrealistic assumption by using a verysmall number of anomalies of confined variability merely to initiate the searchof hyperparameters of the model. We evaluate our solution on natural imagedatasets with a known benchmark, as well as on two medical datasets containingradiology and digital pathology images. The proposed approach suggests a newstrong baseline for image anomaly detection and outperforms state-of-the-artapproaches in complex medical image analysis tasks.


Quick Read (beta)

loading the full paper ...