Memorizing Structure-Texture Correspondence for Image Anomaly Detection

Zhou, Kang; Li, Jing; Xiao, Yuting; Yang, Jianlong; Cheng, Jun; Liu, Wen; Luo, Weixin; Liu, Jiang; Gao, Shenghua

doi:10.1109/tnnls.2021.3101403

Cited by 37 publications

(11 citation statements)

References 60 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Among all the methods, OCGAN [9] has an advantage in the highest computational efficiency (51.3 fps) but takes up relatively large GPU memory (3929MB). The most state-of-the-art method MemSTC [48] consumes more GPU memory (7629MB) and suffers from a relatively low computational efficiency (10.9 fps). SSM reaches 23.5 fps and takes only 2967MB of GPU memory.…”

Section: B Comparison With State-of-the-art Methodsmentioning

confidence: 99%

“…Recently, a deep autoencoder is adopted to improve the feature representation abilities [45]- [48]. In the same vein, MemAE [16] augments the autoencoder with a memory module to highlight reconstructed errors on anomalies.…”

Section: A Unsupervised Anomaly Detectionmentioning

confidence: 99%

“…To capture the information of image texture and structure, P-Net [47] proposes to leverage the relation between the image texture and structure to enlarge the structure difference, which can also be used as a metric for normality measurement. Based on [47], MemSTC [48] proposes a structure-texture correspondence memory module to reconstruct image texture from its structure, where a memory mechanism is used to characterize the mapping from the normal structure to its corresponding normal texture.…”

Section: A Unsupervised Anomaly Detectionmentioning

confidence: 99%

“…For the one-class classification-based method, we consider DSVDD [13] as the baseline. For the reconstruction-based mathods, DAE [42], AE [45], MemAE [16], SCADN [22] and MemSTC [48] are considered. For GAN-based methods, Pix2Pix [75], AnoGAN [41], C-GAN [77], V-GAN [1], EGBAD [49], ALOCC [8], GANomaly [50], f-AnoGAN [74], OCGAN [9] and EnGAN [76] are considered.…”

Section: Mvtec Anomaly Detection Datasetmentioning

confidence: 99%

“…Fig. 8: Qualitative results of anomaly localization of SSM on the MVTec AD dataset for several difficult cases, compared with several state-of-the-art methods, including MKD [63], GP [62], and MemSTC [48]. For SSM, S 0 is the anomaly score map at the initialization step.…”

Section: Ablation Studiesmentioning

confidence: 99%

See 4 more Smart Citations

Self-Supervised Masking for Unsupervised Anomaly Detection and Localization

Huang¹,

Xu²,

Wang³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recently, anomaly detection and localization in multimedia data have received significant attention among the machine learning community. In real-world applications such as medical diagnosis and industrial defect detection, anomalies only present in a fraction of the images. To extend the reconstructionbased anomaly detection architecture to the localized anomalies, we propose a self-supervised learning approach through random masking and then restoring, named Self-Supervised Masking (SSM) for unsupervised anomaly detection and localization. SSM not only enhances the training of the inpainting network but also leads to great improvement in the efficiency of mask prediction at inference. Through random masking, each image is augmented into a diverse set of training triplets, thus enabling the autoencoder to learn to reconstruct with masks of various sizes and shapes during training. To improve the efficiency and effectiveness of anomaly detection and localization at inference, we propose a novel progressive mask refinement approach that progressively uncovers the normal regions and finally locates the anomalous regions. The proposed SSM method outperforms several state-of-the-arts for both anomaly detection and anomaly localization, achieving 98.3% AUC on Retinal-OCT and 93.9% AUC on MVTec AD, respectively.

show abstract

Section: B Comparison With State-of-the-art Methodsmentioning

confidence: 99%