High-Performance Fine Defect Detection in Artificial Leather Using Dual Feature Pool Object Detection

Lin Huang, Weisheng Li, Linlin Shen, Xue Xiao, Suihan Xiao

Abstract

In this study, the structural problems of the YOLOv5 model were analyzed emphatically. Based on the characteristics of fine defects in artificial leather, four innovative structures, namely DFP, IFF, AMP, and EOS, were designed. These advancements led to the proposal of a high-performance artificial leather fine defect detection model named YOLOD. YOLOD demonstrated outstanding performance on the artificial leather defect dataset, achieving an impressive increase of 11.7% - 13.5% in $AP_{50}$ compared to YOLOv5, along with a significant reduction of 5.2% - 7.2% in the error detection rate. Moreover, YOLOD also exhibited remarkable performance on the general MS-COCO dataset, with an increase of 0.4% - 2.6% in AP compared to YOLOv5, and a rise of 2.5% - 4.1% in $AP_S$ compared to YOLOv5. These results demonstrate the superiority of YOLOD in both artificial leather defect detection and general object detection tasks, making it a highly efficient and effective model for real-world applications.

Introduction

Detecting "fine defects" such as tiny spots, stains, or burns in artificial leather is a critical yet challenging task in manufacturing. Traditionally, this process relies on human inspectors, which introduces subjectivity and fatigue. While deep learning models like YOLOv5 have revolutionized object detection, they often struggle with the unique constraints of leather inspection: the microscopic scale of defects relative to the material sheet, and the natural grain "noise" that leads to frequent false positives.

The YOLOD Architecture

To address these challenges, the researchers developed YOLOD, introducing four architectural innovations designed specifically for fine defect detection:

  • Dual Feature Pool (DFP): By "widening" the network's neck, DFP creates specialized feature pools that provide detection heads with richer dimensional data, overcoming the shallow feature mixing of standard FPN+PAN structures.
  • Interference Feature Filtering (IFF): This module acts as a quality control layer, filtering out low-quality feature channels (bottom 0.5% to 5%). This effectively removes background noise from the leather grain that typically confuses detection models.
  • Adaptive Multi Positives (AMP): AMP optimizes the assignment of positive samples during training, ensuring the network learns more effectively from the tiny bounding boxes characteristic of leather defects.
  • Eliminate Output Sensitivity (EOS): By modifying the sigmoid activation function's slope and range, EOS improves model convergence for defects located near the edges of grid cells.

Datasets and Evaluation

The model was evaluated on two primary datasets:

  1. Artificial Leather Dataset (ALD): A custom industrial dataset of 12,000 high-resolution images featuring six defect categories: white spots, stains, particles, black spots, insect marks, and burns.
  2. MS-COCO: The industry-standard benchmark used to verify the model's generalizability across diverse object types.
Visual examples of fine defects in the Artificial Leather Dataset (ALD), categorized into white and black spots groups.

Experimental Results

YOLOD demonstrated significant improvements over baseline models. On the ALD, YOLOD-S achieved a 13.5% increase in $AP_{50}$ compared to YOLOv5-S. Crucially for industrial applications, it also reduced the Average Error (AE)—a measure of false alarms—by up to 7.2%, making it much more reliable for real-world factory floors.

Performance comparison of YOLOD vs. baselines across MS-COCO (AP and small object AP) and the Artificial Leather Dataset (AP50 and Average Error).

On the general MS-COCO dataset, YOLOD-L outperformed YOLOX-L and YOLOv4-CSP in the small object precision ($AP_S$) category by 2.5% and 4.1%, respectively, proving its effectiveness beyond specialized industrial tasks.

Conclusion

YOLOD represents a significant step forward in industrial quality control. By focusing on the "neck" of the detection network and introducing noise-filtering mechanisms, it achieves high precision and low false-alarm rates. Future work will explore the model's applicability to other materials like textiles and metals, as well as further speed optimizations for real-time deployment.

Reference: Huang, L., Li, W., Tan, Y., Shen, L., & Yu, J. (2024). High-Performance Fine Defect Detection in Artificial Leather Using Dual Feature Pool Object Detection. arXiv:2307.16751v3.

Summary updated

Abstract

In this study, the structural problems of the YOLOv5 model were analyzed emphatically. Based on the characteristics of fine defects in artificial leather, four innovative structures, namely DFP, IFF, AMP, and EOS, were designed. These advancements led to the proposal of a high-performance artificial leather fine defect detection model named YOLOD. YOLOD demonstrated outstanding performance on the artificial leather defect dataset, achieving an impressive increase of 11.7% - 13.5% in AP_50 compared to YOLOv5, along with a significant reduction of 5.2% - 7.2% in the error detection rate. Moreover, YOLOD also exhibited remarkable performance on the general MS-COCO dataset, with an increase of 0.4% - 2.6% in AP compared to YOLOv5, and a rise of 2.5% - 4.1% in AP_S compared to YOLOv5. These results demonstrate the superiority of YOLOD in both artificial leather defect detection and general object detection tasks, making it a highly efficient and effective model for real-world applications.
BibTeX
@misc{huang2023high,
  author = {Lin Huang and Weisheng Li and Linlin Shen and Xue Xiao and Suihan Xiao},
  title = {{High-Performance Fine Defect Detection in Artificial Leather Using Dual Feature Pool Object Detection}},
  year = {2023},
  eprint = {2307.16751},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.16751},
}
Figure