Article Details

Abstract

Paper defect detection is a key step in ensuring product quality in the industrial production of paper. Traditional paper defect detection methods rely on manual visual inspection and simple machine vision, which suffer from low detection accuracy, large model size and poor ability to capture small targets. Therefore, this paper proposes a paper defect detection algorithm based on improved YOLOv5. Firstly, the cosine annealing strategy is adopted for periodic dynamic adjustment of the learning rate, which effectively improves the convergence speed and convergence accuracy of the model. Secondly, the ECA attention mechanism is embedded into the backbone network to strengthen the network's ability to focus on and extract key features of paper defects, and improve the detection effect of small targets and weak texture defects. Finally, the optimizer is improved to further enhance the overall performance of the algorithm. The proposed algorithm achieves an mAP@0.5 of 94.39%, an F1-Score of 0.88 and a Recall of 80.56%, with a model size of only 27.1 MB. Compared with the original YOLOv5, its mAP@0.5, Recall and F1-Score are increased by 5.18 percentage points, 18.75 percentage points and 0.12 respectively. Compared with the model without the ECA attention mechanism, its mAP@0.5 and Recall are improved by 0.59 percentage points and 0.35 percentage points. Compared with the SGD optimizer, its mAP@0.5, Recall and F1-Score are raised by 3.83 percentage points, 14.59 percentage points and 0.09 respectively. Experimental results show that the improved YOLOv5-based paper defect detection model can provide a reference for subsequent related research.

Keywords

defect detection; ECA attention; cosine annealing; optimizer

References

1. [1] Party Jiaqi, Wang Wanxing Research on Optimization of Paper Defect Detection System by Deep Learning Algorithm from the Perspective of Intelligence [J]. Paper Science and Technology, 2025, 44 (9): 96-99
2. [2] Chen Yumei, Li Zhaofei, Hou Jin, etc Research on Improving Paper Surface Defect Detection of YOLOv7 [J]. China Paper Industry, 2023, 42 (11): 160-168
3. [3] Zhou Qiang, Chen Ying, Shen Tianyu, etc A review of the development of paper disease detection systems based on machine vision [J]. China Paper Industry, 2016, 35 (5): 72-79
4. [4] Xu Jilong, Zhao Shuai, Wu Dandan, etc Fabric surface defect recognition algorithm based on lightweight improvement YOLOv5s [J]. Synthetic Fiber, 2026, 55 (2): 49-56
5. [5] Zhao Biao, Li Wenling, Zhu Chengguang Research on paper defect detection network improved by YOLOv5 based on backbone network and feature fusion [J]. Paper Science and Technology, 2025, 44 (12): 116-118
6. [6] Zhang Yue, Zhang Peng, Ma Qinghua A fast detection method for occluded faces in natural scenes based on improved YOLOv3 [J]. Computer and Digital Engineering, 2025, 53 (12): 3507-3512
7. [7] Shen Siyi, Xie Mingying Research on Defect Detection of Cigarette Trademark Paper Based on Deep Learning and Machine Vision [J]. Printing Industry, 2025 (4): 44-48
8. [8] Jia Yaohuan, Huang Kunming, Yang Yuchao, etc Intelligent recognition of mung bean leaf diseases based on improved EfficientNet and transfer learning [J/OL]. Agricultural Mechanization Research (2026-02-26) [2026-04-01] https://link.cnki.net/urlid/23.1233.S.20260225.2345.020.
9. [9] Guo Wenqi, Yang Guowei, Huang Luyao, etc Electronic component detection and classification based on YOLOv8 and improved ResNet50 [J]. Journal of Tianjin University of Science and Technology, 2026, 41 (1): 61-68
10. [10] Luo Hao, Li Xuebing, Gu Yujian Traffic sign detection method based on improved YOLOv5s [J]. Computer and Digital Engineering, 2025, 53 (12): 3513-3518
11. [11] Yao Senshan, Pang Chengxin Data anomaly detection method for photovoltaic inverters based on improved residual network [J]. Journal of Solar Energy, 2025, 46 (4): 322-330
12. [12] Chen Yongqiang Research and Application of Lightweight Facial Expression Recognition Algorithm Based on Deep Learning [D]. Nanjing: Nanjing University of Information Science and Technology, 2025
13. [13] Song Lei Trans YOLO: Defect Detection Method for Photovoltaic Modules Based on Deep Learning [J]. Power Plant Systems Engineering, 2025, 41 (6): 13-16
14. [14] He Wanlong, Geng Hongwei, Zhang Feifei, etc Research on Image Recognition System for Important Wheat Diseases Based on Deep Learning [J]. Journal of Crop Science, 2026, 52 (5): 1401-1417
15. [15] Sun Haijing, Cui Jiaqi, Shao Yichuan, etc Gastrointestinal endoscopy image classification method based on improved Adam algorithm [J]. Journal of Shenyang University (Natural Science Edition), 2026, 38 (1): 53-60, 90
16. [16] Li Guo A paper defect automatic detection method based on Faster RCNN improved algorithm [J]. Journal of Gannan Normal University, 2021, 42 (3): 100-104

Funding

Project of Baoji Vocational and Technical College (2024250Y)

About the authors

Wen Zhaofa (1996-), male, Han ethnicity, from Baoji, Shaanxi Province, teaching assistant, master's degree, research direction: computer vision.