Evaluation of Nested U-Net models performance on MVTec AD dataset

Anomaly detection (AD) from image data using convolutional neural networks and deep learning has become a widespread topic among both scientists and engineers. In addition to the development of new methods and models, specialized datasets are created as well. The most cited dataset specialised on anomaly detection tasks and created for testing the most recent methods is MVTec AD. This dataset has been used in more than 40 articles which are mainly devoted to creating or modifying AD methods. Subsequently, their performance is usually tested on the MVTec AD dataset. However, despite a large number of different methods and models, there is a lack of performance evaluation of U-Net++ (Nested U-Net architecture), a robust model which is well-known in the field of segmentation tasks. This article is focused on the evaluation of two Nested U-Net architectures (U-Net++, ANU-Net) on the MVTec AD dataset. It is shown that the direct use of the Nested U-Net models to reconstruct anomaly-free input data together with their strong augmentation during training phase leads to inability to reconstruct image data with anomalies at inference time. Achieved results can compete with some of the state-of-the-art reconstruction-based methods. The average image-level AUROC performance of U-Net++ model is 97.9% and 96.2% for image size of
$64\times 64$
and
$128\times 128$
pixels, respectively. Further, the average performance of ANU-Net on image-level detection is 96.5% and 96.8% for image size of
$64\times 64$
and
$128\times 128$
pixels, respectively.