Fabric Defect Detection Based on Improved YOLOv7 Network
Abstract
In order to achieve real-time defect detection technology, detection accuracy, prediction speed, and lightweight deployment models
are important issues. Traditional object detection methods often fail to achieve a balanced effect on all aspects. Therefore, a detection model based on lightweight convolutional neural network YOLOv7 is proposed. Firstly, lightweight convolutional Ghost conv is introduced to
lighten the backbone network; Secondly, adding CBAM attention mechanism to suppress invalid information and enhance feature extraction
ability; Finally, a new measurement method is introduced at the regression loss function α- Replacing IoU with SIoU accelerates algorithm
convergence and improves detection efficiency for defect targets. The experiment shows that the accuracy P of the detection model reaches
96.27%, the mAP index is 83.84%, the detection speed is 23.83ms, and the model size is only 19.10MB, effectively balancing the accuracy,
real-time performance, and lightweight deployment of defect detection.
are important issues. Traditional object detection methods often fail to achieve a balanced effect on all aspects. Therefore, a detection model based on lightweight convolutional neural network YOLOv7 is proposed. Firstly, lightweight convolutional Ghost conv is introduced to
lighten the backbone network; Secondly, adding CBAM attention mechanism to suppress invalid information and enhance feature extraction
ability; Finally, a new measurement method is introduced at the regression loss function α- Replacing IoU with SIoU accelerates algorithm
convergence and improves detection efficiency for defect targets. The experiment shows that the accuracy P of the detection model reaches
96.27%, the mAP index is 83.84%, the detection speed is 23.83ms, and the model size is only 19.10MB, effectively balancing the accuracy,
real-time performance, and lightweight deployment of defect detection.
Keywords
effect detection; YOLOv7; CBAM Attention Module; Ghost Convolution; -SIoU
Full Text:
PDFReferences
[1] Zeng Huafu, Yang Jie, Li Linhong. Clothing image classification algorithm based on improved ShuffleNet v1 [J]. Modern Textile
Technology, 2019,31(2):23-35. (in Chinese) DOI:10.19398 / j.AT.202208003.
[2] Zhang Chun-chun, Luo Ruilin, Lu Lin, et al. Image recognition of microstrip green tobacco leaf based on MobileNet and transfer
learning [J]. Journal of Yunnan Normal University (Natural Science Edition), 2023,43(4):46-48. DOI:10.7699/j.nnu. ns-2023-052.
[3] Zhang Qingchun, Wu Zheng, Zhou Ling, et al. Vehicle and Pedestrian target recognition method based on improved YOLOv5 [J].
China Tests and Tests,2023,49(7):82-88. (in Chinese) DOI:10.11857/j.issn.1674-5124.2022060008.
DOI: http://dx.doi.org/10.18686/ag.v8i1.12588
Refbacks
- There are currently no refbacks.
