TruingDet: Towards high-quality visual automatic defect inspection for mental surface. (March 2021)

• Record Type: Journal Article
• Authors: Liu, Zhenyu; Tang, Ruining; Duan, Guifang; Tan, Jianrong
• Is Part Of: Optics and lasers in engineering. Volume 138(2021)
• Publication Date: 2021-03
• Subjects: Automatic defect detection -- Balanced feature pyramid -- Cascade head -- Deep learning -- Deformable convolution -- High quality defect detection; Lasers in engineering -- Periodicals; Optical measurements -- Periodicals; Optics -- Periodicals; Lasers en ingénierie -- Périodiques; Mesures optiques -- Périodiques; Optique -- Périodiques; 621.36605
• DOI: 10.1016/j.optlaseng.2020.106423
• Languages: English
• ISSNs: 0143-8166

Abstract

Highlight: Deformable convolution and balanced feature pyramid are introduced to address the irregular variations in shape and scale variations respectively. High-quality localization result is obtained based on the cascade head module. The proposed model achieves state-of-the-art (SOTA) results on COCO mean Average Precision (mAP) metrics compared with baseline and other popular detectors. Abstract: Visual surface defect detection, which aims to obtain the locations of defects and classify each defect into the corresponding category in a given image, is a critical task in an actual production process. Nowadays, more and more methods have made excellent progress in visual defect inspection. However, there still exist three tough challenges where these methods cannot handle well: large defect shape change, large-scale variation, and high-quality defect localization. In this paper, a Convolutional Neural Networks (CNN) based visual defect detection framework is proposed, which elegantly mitigated these three problems by introducing three well-designed components including deformable convolution module, balanced feature pyramid module and cascade head module. First, the feature maps contained with defect shape information are adaptively extracted by Resnet/ResneXt network with the deformable convolution operator. Then the balanced feature pyramid module is attached to the feature extraction module to obtain information-fused multilayer feature maps. Finally, the cascade head is applied to refine the predicted bounding box to achieve high-quality defect localization. Under the COCO evaluation metrics, our method significantly obtains 45.2 mAP with a large margin (4.9 AP) compared with Faster RCNN baseline.