Improved YOLOv11-Based Steel Pipe Inner Wall Defect Detection Method

Literature Overview

This paper by Qi Xiangjing, Gao Hongwei, and Zhang Xihan from Shenyang Ligong University, published in Journal of Shenyang Ligong University (2025, Vol. 44, Issue 6, pp. 7-13), presents an improved data analysis-based defect detection model for steel pipe inner wall inspection. The proposed model, based on YOLOv11 architecture, incorporates Slim-neck modules and Efficient Multi-Scale Attention (EMA) mechanisms to address challenges in industrial defect detection including variable illumination, texture interference, and low contrast between defects and backgrounds. The work is funded by the Liaoning Provincial Department of Education (Grant No. LJ232410144070).

Problem Statement and Motivation

Steel pipe inner wall defect detection is critical for ensuring product quality and production safety. Traditional inspection methods—including visual inspection, eddy current testing, and magnetic flux leakage methods—face significant limitations:

• Visual inspection: Subject to operator fatigue, limited by lighting conditions and spatial constraints inside the pipe.
• Eddy current testing: Requires close contact or proximity to the surface, sensitive to lift-off effects, and may miss certain defect types.
• Magnetic flux leakage: Primarily effective for ferromagnetic materials and limited in defect characterization capability.

The paper identifies that while data analysis approaches have shown promise in defect detection, they still struggle with illumination variations, texture interference, and the challenge of distinguishing defects from background features with similar visual characteristics.

Technical Methodology

YOLOv11 Base Architecture

YOLOv11 (You Only Look Once version 11) is a single-stage object detection framework that processes images in a single pass, offering real-time detection capability. The base model consists of a backbone for feature extraction, a neck for feature fusion, and a head for detection output.

Slim-Neck Module Enhancement

The proposed improvement introduces Slim-neck modules that effectively fuse features extracted from standard convolutions and depthwise separable convolutions. The key advantages of this approach include:

• Reduced computational complexity: Depthwise separable convolutions significantly reduce the number of parameters compared to standard convolutions.
• Enhanced feature extraction: By combining features from both convolution types, the model captures both spatial and channel-wise information more comprehensively.
• Efficient multi-scale representation: The fusion mechanism preserves information across different scale levels, improving detection of defects of varying sizes.

Efficient Multi-Scale Attention (EMA) Mechanism

The EMA mechanism is integrated to further enhance feature representation. Its key design features include:

• Parallel sub-structure design: Reduces sequential dependencies in the computation graph, improving inference speed.
• Multi-scale attention: Captures contextual information at multiple scales, which is particularly beneficial for detecting defects of varying sizes.
• Efficient parameter utilization: Achieves attention-based feature enhancement without excessive parameter overhead.

Performance Results

The 11.8% improvement in mean average precision (mAP@0.5) on the self-built dataset demonstrates the effectiveness of the proposed modifications. The simultaneous reduction in parameters and computational cost indicates that the improvements are not achieved at the expense of efficiency—a critical consideration for industrial deployment where real-time detection is required.

Engineering Practice Implications

Industrial Deployment Considerations

For industrial implementation, several factors beyond accuracy must be considered:

1. Lighting robustness: The model must perform consistently across varying illumination conditions encountered in production environments.
2. Real-time capability: Detection speed must be compatible with production line speeds, typically requiring processing times under 100 milliseconds per frame.
3. Edge deployment: Industrial settings often require on-premise processing rather than cloud-based solutions, necessitating optimization for edge computing hardware.
4. Continuous learning: The model should be periodically retrained with new defect samples to maintain performance as production conditions evolve.

Integration with Existing Inspection Systems

The proposed method can complement traditional inspection methods rather than replace them entirely. A hybrid approach combining ultrasonic testing, eddy current testing, and visual inspection with data analysis-based image analysis can provide comprehensive defect coverage. The data analysis component is particularly valuable for defect classification, sizing, and prioritization, which traditional methods often cannot provide efficiently.

Critical Reflections and Questions

The paper reports results on a self-built dataset, which raises questions about generalizability. Defect detection models trained on one manufacturer's pipes may not perform well on pipes from different manufacturers due to differences in surface finish, material composition, and typical defect patterns. Cross-validation across multiple data sources would strengthen the claims of the paper.

Additionally, the paper does not extensively discuss the types of defects detected. Steel pipe inner wall defects can include scratches, dents, inclusions, cracks, pitting corrosion, and manufacturing residues. The relative performance of the model on different defect types would be valuable information for potential users.

From a practical standpoint, the paper could have discussed the data annotation process in more detail. The quality and consistency of defect annotations significantly affect model performance, and the annotation methodology should be described to enable replication by other researchers.

Study Insights and Reference Value

This paper represents a meaningful contribution to the application of data analysis in steel pipe quality inspection. The combination of Slim-neck modules and EMA mechanisms offers a practical path to improved detection accuracy without excessive computational overhead. For steel pipe manufacturers and quality assurance engineers, this approach suggests a viable pathway toward automated, real-time inner wall defect detection that can supplement or partially replace manual visual inspection. The key challenge moving forward will be adapting such models to diverse production environments and ensuring consistent performance across varying operating conditions.