Deep industrial image anomaly detection: A survey

Researchers review the recent advances of deep learning-basedimage anomaly detection since the rapid development ofdeep learning can bring the capabilities of image anomaly detection into the factory floor. In modern manufacturing, image anomaly detection (IAD) is always performed at the end of the manufacturing process and tries to identify product defects. The price of a product is significantly affected by the defect’s severity. In addition, if the flaw reaches a certain threshold, the product will bediscarded. Historically, most anomaly detection tasks are performed by humans, which suffers fromthe following many disadvantages: Firstly, it is impossible to avoid human fatigue, resulting in a false positive phenomenon (i.e., the ground truth is abnormal, while the human’s judgment is normal). Secondly, long and intensive work on anomaly detection maycause health problems, such as visual impairment. Thirdly, locating anomalies requires a significant number ofemployees, raising operational costs.

 

Thus, the goal of IAD algorithms is to reduce humanlabor and improve productivity and product quality. Before deep learning, the performance of IAD could not fulfil the demands of industrial manufacturing. Nowadays, the deep learning method has received good results, andmost of these methods are more than 97% accurate. Still,IAD has many problems when it comes to real-world use.To comprehensively explore the effectiveness and applicable scenarios of the current methods, more careful analysis of IAD researchers conduct in this survey is necessary andsignificant.

 

As a representative review that focusesmore on traditional methods, Czimmermann et al. haveless discussion of deep learning methods, while this survey discusses deep learning in more depth. Firstly, this study uses twice as many IAD datasets as Tao et al. Secondly, researchers analyze the performance of IAD using themost comprehensive image level and pixel level metrics.Nevertheless, Cui et al. and Tao et al. only employimage level metrics, neglecting the anomalies localizationperformance of IAD. Thirdly, this study develops a taxonomy based on the design of neural network architecture with varying degrees of supervision. Finally, tobridge the gap between academic research and real-worldindustry needs, researchers review the current IAD algorithms under industrial manufacturing settings.

 

As an emerging field, research on IAD must fully consider industrial manufacturing requirements. The following is a summary of the challenging issues that need to beinvestigated: 1) IAD dataset should be gathered from actual manufacturing lines, not labs. The public cannot access thereal-world anomalous dataset due to privacy concerns.The majority of open-source IAD datasets generate anomalies from anomaly-free products. In other words, the abnormalities from open-source IAD datasets may not occur in actual production lines, which makes deployingIADs in industrial manufacturing very challenging. 2) It is challenging to enable the creation of a unifiedIAD model in the absence of multiple domain IAD datasets. Recently, You et al. propose a unified IAD modelfor multiple class objects. However, they disregard thenotion that commodities produced in the same plantshould be of the same sort. For example, an automakermanufactures several types of workpieces but does notproduce fruit. Current popular IAD datasets, like MVTecAD and MVTec LOCO, consist of numerous classesbut not multiple domains. To simulate a realistic manufacturing process, researchers must create a new IAD dataset collected from multiple domains. 3) It is urgent to set up a uniform assessment for theimage-level and pixel level of IAD performance. The majority of IAD metrics shrink the anomalous mask (groundtruth) into the size of feature map for evaluation, whichinevitably reduces the precision of assessment. Moreover, researchers discover that certain IAD methods perform well onimage AUROC but poorly on pixel AP, or vice versa.Therefore, it is essential to develop a uniform metric forassessment IAD performance at both image and pixellevel. 4) Researchers should design a more efficient loss function thatcan leverage both the guidance of labelled data and theexploration of unlabeled data. In realistic manufacturingscenario, limited number of anomalous samples are available. However, most of unsupervised IAD methods outperform semi-supervised IAD methods. By observing thefailure of semi-supervised IAD, researchers would call for more attention to the feature extraction and loss function, whichcan leverage both the guidance from labels efficiently andthe exploration from the unlabeled data. Regarding thekey problem mentioned above, improving feature extraction from abnormal samples and redesigning deviationloss function can fully use labelled anomalies and divergethe feature space of abnormal samples from those of normal samples.

 

The paper categorizes various methods into severalparadigms, and clearly analyzes the advantages and disadvantages of various paradigms. It allows the reader tounderstand the state-of-the-art quickly and provides a reliable guide for selecting the required algorithm for practical applications. More importantly, researchersinthispaper have analyzedthe disadvantages of different paradigms and the currentmain challenges. Subsequent researchers can quickly finddirections to push the field forward.

 

The main contributions of this survey can be summarized as following:1) Researchers provide an in-depth review of image anomalydetection by considering the design of neural network architecture with varying degrees of supervision.2) It provides a comprehensive review of the currentIAD algorithms in different settings to bridge the gapbetween the academic research and real-world industrialmanufacturing.3) It summarizes the main issues and potential challenges in IAD, which outlines the underlying research directions for future works.

 

Section 2 isaboutunsupervised anomaly detection. Most current research focuses on unsupervised anomaly detection, based on the assumption thatthe collection of abnormal samples incurs massive humanand financial costs. This indicates that only normalsamples are included in the training set, whereas both abnormal and normal samples are included in the test set.Anomaly detection in industrial images is a subset ofproblems with out-of-distribution (OOD). Before the riseof deep learning, differential detection and filtering werefrequently used to detect anomalies in industrial images.Following the release of the MVTec AD, methods foranomaly detection in industrial images can be divided into two categories: feature-embedding and reconstructed-based. Currently, more AD techniques are based on feature embedding.

 

Section 3 introducessupervised anomaly detection. Despite the fact that abnormal data is diverse and difficult to collect, it is still possible to collect abnormalsamples in real-world scenarios. Therefore, some researches focus on how to train models for anomaly detection using a small number of abnormal samples and many normal samples. Besides, a number of studies fail to account for theunbalanced distribution of normal and abnormal samplesand rely primarily on abnormal samples for supervisedtraining. This part providesmanyexamples as proof.

 

Section 4 introduces the classification standards orapplication settings that are more appropriate for industrial scenes, namely few-shot anomaly detection, noisy anomaly detection, anomaly synthesis, and 3D anomaly detection.

 

Researchers describe the popular dataset in Section 5 and take a retrospective view of the metrics function in Section 5. Thereare two Tablesinthispart, one demonstrates that the number and the sizeof IAD dataset are gradually increasing, but most of themare not generated in a real production line; the other one offers a comprehensive review of themetrics in industrial image anomaly detection.

 

Then, researchers provide an analysis of the performance of current IAD methods on various datasets in Section 6. Thereare two Tables in this part that show the statistical result of currentIAD performance on MVTec AD. Finally, researchers provide future research directions for IAD in Section 7.

 

To conclude, in this paper, researchers provide a literature review on imageanomaly detection in industrial manufacturing, focusingon the level of supervision, the design of neural networkarchitecture, the types and properties of datasets and theevaluation metrics. In particular, they characterize thepromising setting from industrial manufacturing and review current IAD algorithms in their proposed setting. Inaddition, researchers investigate in depth which network architecture design can considerably improve anomaly detectionperformance. In the end, they highlight several exciting future research directions for image anomaly detection.

 

See the article:

Deep Industrial Image Anomaly Detection: A Survey

http://doi.org/10.1007/s11633-023-1459-z

---

---