Same Same But DifferNet: Semi-Supervised Defect Detection with Normalizing Flows

Abstract

The detection of manufacturing errors is crucial in fabrication processes to ensure product quality and safety standards. Since many defects occur very rarely and their characteristics are mostly unknown a priori, their detection is still an open research question. To this end, we propose Dif-ferNet: It leverages the descriptiveness of features extracted by convolutional neural networks to estimate their density using normalizing flows. Normalizing flows are well-suited to deal with low dimensional data distributions. However, they struggle with the high dimensionality of images. Therefore, we employ a multi-scale feature extractor which enables the normalizing flow to assign meaningful likelihoods to the images. Based on these likelihoods we develop a scoring function that indicates defects. Moreover, propagating the score back to the image enables pixel-wise localization. To achieve a high robustness and performance we exploit multiple transformations in training and evaluation. In contrast to most other methods, ours does not require a large number of training samples and performs well with as low as 16 images. We demonstrate the superior performance over existing approaches on the challenging and newly proposed MVTec AD [4] and Magnetic Tile Defects [14] datasets.

CutPaste: Self-Supervised Learning for Anomaly Detection and Localization

“…Because of practical applications, such as industrial inspection or medical diagnosis, defect detection [9,5] has received lots of attention. The initial steps have been taken with methods including autoencoding [9,7,25,59], generative adversarial networks [48,3], using pretrained models on ImageNet [38,45,6,14,43,44], and self-supervised learning by solving different proxy tasks with augmentations [61,47,57,15]. The proposed CutPaste prediction task is not only shown to have strong performance on defect detection, but also amenable to combine with existing methods, such as transfer learning from pretrained models for better performance or patch-based models for more accurate localization, which we demonstrate in Section 4.…”

Piksel seviyesinde yüzey hata tespiti için InceptionV3 tabanlı zenginleştirilmiş öznitelik entegrasyon ağ mimarisi

“…Bu yöntemlere ek olarak literatürde Auto-Encoder (Oto-Kodlayıcı) tabanlı denetimsiz yüzey hatası algılama yöntemleri geliştirilmiştir. DifferNet [29] modelinde, önceden eğitilmiş AlexNet mimarisinden çok ölçekli bir özellik çıkarıcı modülü ile anlamlı öznitelikler çıkarılmıştır. Tahmini bir hata konumu elde etmek için bu özelliklere Normalleştirme Akışı yöntemi kullanılmıştır.…”

AdeNet: Deep learning architecture that identifies damaged electrical insulators in power lines

“…In [3], a comprehensive real-world dataset is curated for unsupervised detection of anomalies. A common approach is to train the classifier solely on the normal samples with the hope that the architecture will be robust enough to capture the intrinsic characteristic features of the normal class and thus can identify abnormal characteristics by inference [23]. Notable state-of-the-art methods include AnoGAN [26,25], L2 and SSIM Autoencoder [4,3], CNN Feature Dictionary [20], GMM-Based Texture Inspection Model [5], and Variation Autoencoder.…”