Anomaly Detection Using GANs for Visual Inspection in Noisy Training Data

Abstract

The detection and the quantification of anomalies in image data are critical tasks in industrial scenes such as detecting micro scratches on product. In recent years, due to the difficulty of defining anomalies and the limit of correcting their labels, research on unsupervised anomaly detection using generative models has attracted attention. Generally, in those studies, only normal images are used for training to model the distribution of normal images. The model measures the anomalies in the target images by reproducing the most similar images and scoring image patches indicating their fit to the learned distribution. This approach is based on a strong presumption; the trained model should not be able to generate abnormal images. However, in reality, the model can generate abnormal images mainly due to noisy normal data which include small abnormal pixels, and such noise severely affects the accuracy of the model. Therefore, we propose a novel anomaly detection method to distort the distribution of the model with existing abnormal images. The proposed method detects pixel-level micro anomalies with a high accuracy from 1024×1024 high resolution images which are actually used in an industrial scene. In this paper, we share experimental results on open datasets, due to the confidentiality of the data.

Learning Competitive and Discriminative Reconstructions for Anomaly Detection

“…We use similar setup as in (You, Robinson, and Vidal 2017), the experimental results are presented in Table 3. The results of the other methods are from (Sabokrou et al 2018) and (Kimura and Yanagihara 2018). These results show that even for a small number of training samples, our method performs at least as well as the state-of-the-art algorithms, and in many cases it superior to them.…”

“…Considering that Caltech-256 is a benchmark dataset for outlier detection task, in addition to DAOC, we compare our method with 7 other methods therein designed specifically for detecting outliers. Those methods include Coheerence Pursuit (CoP) (Rahmani and Atia 2017), OutlierPursuit (Xu, Caramanis, and Sanghavi 2010), REAPER(Lerman et al 2015), Dual Principal Component Pursuit (DPCP) (Tsakiris and Vidal 2015), Low-Rank Representation (LRR) (Liu, Lin, and Yu 2010), OutRank (Moonesignhe and Tan 2006), and inductive semi-supervised GAN (SSGAN) (Kimura and Yanagihara 2018).…”

“…Note that our formulation for anomaly detection is different from positive and unlabeled (PU) learning (Elkan and Noto 2008) or inductive semi-supervised anomaly detection (Kimura and Yanagihara 2018). In both cases, it is assumed that the unlabeled data that used for training share the same distribution with the test data.…”

Anomaly Detection Based on Zero-Shot Outlier Synthesis and Hierarchical Feature Distillation

“…Data Representation. Existing methods rely on low level features, e.g., Histogram of Oriented Gradients [29], local patterns [30]- [34]; high level features, e.g., bag of words [35], trajectories [36]; or deep-learned features [8], [9], [23], [24], [26], [27], [37], [38]. The latter have gained traction in the past years since they tend to outperform low or high-level features designed for general purposes since the deep-features are tuned for each particular task.…”

“…For our experiments on MNIST, Caltech-256, Coil-100, and Yale B, we compared our proposed methods against a set of methods, namely LOF [18], DRAE [10], R and RD [9], GPND [24], Coherent Pursuit (CoP) [20], REAPER [14], Outlier Pursuit (OP) [39], LRR [75], DPCP [76], 1 -thresh. [77], R-Graph [11], AnoGAN [27], and AGAN [23].…”