Zero-DD: Zero-sample defect detection for industrial products. (January 2023)

• Record Type: Journal Article
• Authors: Zhang, Haigang; Pan, Ronghui; Chang, Fengjun; He, Liming; Dong, Zhongming; Yang, Jinfeng
• Is Part Of: Computers & electrical engineering. Volume 105(2023)
• Publication Date: 2023-01
• Subjects: Industrial defect detection -- Generative adversarial network -- Zero sample detection; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2022.108516
• Languages: English
• ISSNs: 0045-7906

Abstract

Abstract: In current industrial production, defective products have gradually become the bottleneck for enterprises to improve product quality. Artificial detection is far less rapid, accurate and stable than machine vision based detection. However, due to the inevitable emergence of unknown defects, data-driven based supervised recognition algorithms often fail and cannot be practically applied. In this paper, a detection model for surface defects of industrial products with zero samples, termed Zero-DD, is proposed, where three important components can ensure that the model is sensitive to unknown defects and achieve high-precision detection. Firstly, we propose a differential feature extraction module based on the Siamese network, which adopts the dual-stream input channels to strengthen the feature difference. Such operation can better highlight defect features and effectively deal with the feature extraction for defects with small sizes. Secondly, the differential features are fed into the Generative Adversarial Network (GAN) after being coupled with the input of the generator. We argue that GAN model are able to distinguish between normal and defective samples by learning from known samples. When the model encounters unknown defects, it is unable to generate images that satisfy the discriminator to realize the recognition of unknown defects. Thirdly, in order to improve the model detection accuracy, the Channel and Coordinate Attention module is proposed, which realizes attention assistance from the perspective of channels and coordinates. Additionally, in this paper, we publish a novel visual dataset of injection-molded bottle cap defects, called BC defects, which contains 3008 samples with 8 types of defects. Finally, simulation experiments based on BC defects dataset have demonstrated the effectiveness of the proposed Zero-DD model.