Zhuojin Pipe Fitting Co., LtdYOLO-EGVS Algorithm for Recognition of Scattered Stacked Tee Fittings
Literature Overview and Industrial Motivation
Published in Modular Machine Tool & Automatic Machining Technique (2025, No. 7, pp. 45-51), this paper addresses a practical industrial problem: the automated visual recognition of tee fitting parts that are scattered and stacked on a production line. The authors from Taiyuan University of Science and Technology propose an improved object detection algorithm called YOLO-EGVS, based on the YOLOv8n architecture, to achieve high-accuracy detection of tee fittings in cluttered industrial scenes. The motivation is clear — in modern manufacturing environments, automated part recognition is essential for robotic picking, quality inspection, and inventory management, and tee fittings present particular challenges due to their complex geometry, similar shapes among different sizes, and tendency to overlap when randomly stacked.
Algorithm Architecture and Key Modifications
The YOLO-EGVS algorithm introduces three targeted modifications to the YOLOv8n baseline:
| Component | YOLOv8n Baseline | YOLO-EGVS Modification | Purpose |
|---|---|---|---|
| Backbone | CSPDarknet | CSPDarknet + EMA attention | Enhanced feature extraction |
| Neck | C2f modules | GSConv + VOV-GSCPS (SlimNeck) | Reduced complexity, faster detection |
| Detection Head | CIOU loss | Shape-IOU loss | Better bounding box accuracy |
The EMA (Exponential Moving Average) attention mechanism added to the backbone helps the network focus on the most discriminative features of tee fittings, such as the characteristic three-port geometry and flange or threaded connection features. The SlimNeck architecture, incorporating GSConv (Ghost Convolution) and VOV-GSCPS modules, reduces the computational complexity of the neck while maintaining or improving feature fusion quality. The Shape-IOU loss function, which replaces the standard CIOU (Complete IOU) loss, pays additional attention to the shape and scale of the bounding box, which is critical for distinguishing between different sizes of tee fittings.
Performance Results and Comparative Analysis
The experimental results demonstrate that the YOLO-EGVS algorithm outperforms the YOLOv8n baseline across all key metrics:
| Metric | YOLOv8n | YOLO-EGVS | Improvement |
|---|---|---|---|
| Precision (P) | 93.7% | 94.7% | +1.0% |
| Recall (R) | 95.6% | 96.5% | +0.9% |
| mAP | 95.6% | 96.9% | +1.3% |
| Model size | 6.3 MB | 5.9 MB | -6.3% |
| Computational cost | 8.2 GFLOPs | 7.4 GFLOPs | -9.8% |
When compared with other lightweight detection models (YOLOv3-tiny, YOLOv5n, YOLOv6n), the YOLO-EGVS algorithm achieves a superior balance between model complexity and detection performance, making it well-suited for deployment on industrial edge computing hardware.
Engineering Practice Considerations
For manufacturing engineers implementing this technology in a real production environment, several practical aspects deserve attention:
- Lighting conditions: Industrial environments often have variable and harsh lighting. The algorithm should be validated under a range of lighting conditions representative of the actual production floor.
- Occlusion handling: When tee fittings are heavily stacked, significant occlusion occurs. The algorithm's ability to detect partially visible parts should be tested specifically.
- Multi-class discrimination: If multiple sizes and types of tee fittings need to be distinguished, the dataset must include sufficient examples of each class, and the algorithm's class separation capability must be verified.
- Real-time performance: The 7.4 GFLOPs computational cost should be evaluated against the specific hardware platform to ensure real-time detection rates suitable for the production line speed.
- Robustness to surface conditions: Manufacturing residues, oil, paint marks, and handling marks on tee fittings may affect detection accuracy and should be considered in dataset construction.
Study Insights and Reflections
This paper demonstrates a pragmatic approach to industrial computer vision: starting from a proven baseline architecture and making targeted, purposeful modifications rather than designing an entirely new algorithm from scratch. The combination of attention mechanisms for feature enhancement, lightweight modules for efficiency, and improved loss functions for accuracy represents a well-balanced optimization strategy. From a manufacturing engineering perspective, the practical value of this work lies in enabling reliable automated recognition of complex pipe fitting geometries in real-world production scenarios. The relatively small model size (5.9 MB) and low computational cost (7.4 GFLOPs) make it feasible for deployment on embedded industrial hardware, which is a critical practical consideration that many purely academic detection papers overlook.