Iterative energy-based projection on a normal data manifold for anomaly localization

Abstract

Autoencoder reconstructions are widely used for the task of unsupervised anomaly localization. Indeed, an autoencoder trained on normal data is expected to only be able to reconstruct normal features of the data, allowing the segmentation of anomalous pixels in an image via a simple comparison between the image and its autoencoder reconstruction. In practice however, local defects added to a normal image can deteriorate the whole reconstruction, making this segmentation challenging. To tackle the issue, we propose in this paper a new approach for projecting anomalous data on a autoencoder-learned normal data manifold, by using gradient descent on an energy derived from the autoencoder's loss function. This energy can be augmented with regularization terms that model priors on what constitutes the user-defined optimal projection. By iteratively updating the input of the autoencoder, we bypass the loss of high-frequency information caused by the autoencoder bottleneck. This allows to produce images of higher quality than classic reconstructions. Our method achieves state-of-the-art results on various anomaly localization datasets. It also shows promising results at an inpainting task on the CelebA dataset.

TrustMAE: A Noise-Resilient Defect Classification Framework using Memory-Augmented Auto-Encoders with Trust Regions

“…To get a segmentation map, we can directly impose a threshold on the reconstruction errors. Table 2 shows our defect segmentation performance in comparison with several recent works on defect segmentation [3,5,6,12,19,22,23,34]. We can observe that our method consistently performs competitively across all classes, achieving the best mean AUC among all the baselines.…”

“…Techniques for defect detection can be broadly grouped into: classification-based [11,24,32,35,36], detection-based [10,37], segmentation-based [7,10,17,20,28,[38][39][40], and reconstruction-based [3,5,6,23,25,34,41,47].…”

“…We adopted most of the baseline performance values from Bergmann et al [2], Dehaene et al [6], Huang et al [16], and Liu et al [22].…”

“…The baselines AE-SSIM [3] and AE-L2 [2] follows the setting of Dehaene et al [6]. They used the network architecture of Bergmann et al [3] with a latent dimension of 100 and trained for 300 epochs with a learning rate of 1e −4 .…”

Deep One-Class Classification via Interpolated Gaussian Descriptor

“…4 shows anomaly localisation results on MVTec AD images, where red regions in the heatmap indicate higher anomaly probability. From this results, we can see that our approach can localise anomalous regions of different sizes and structures from different object Metric Method Mean Accuracy AVID (Sabokrou et al 2018) 0.730 AESSIM (Bergmann et al 2018) 0.630 DAE (Hadsell et al 2006) 0.710 AnoGAN (Schlegl et al 2017) 0.550 λ-VAEu (Dehaene et al 2020) 0.770 LSA (Abati et al 2019) 0.730 CAVGA-Du (Venkataramanan et al 2019) (Bergmann et al 2018) 0.87 AVID (Sabokrou et al 2018) 0.78 SCADN (Yan et al 2021) 0.75 LSA (Abati et al 2019) 0.79 λ-VAEu (Dehaene et al 2020) 0.86 AnoGAN (Schlegl et al 2017) 0.74 ADVAE (Liu et al 2020) 0.86 CAVGA-Du (Venkataramanan et al 2019) 0.85 CAVGA-Ru (Venkataramanan et al 2019) 0.89 Ours -ImageNet 0.91 Ours -SSL 0.93…”