New Algorithm Enhances Defect Detection in Power Plant Pipelines Using Infrared Imaging
Accurate and timely detection of defects in high‑temperature pipelines is critical to the safe and efficient operation of thermal power plants. Traditional manual inspections using handheld infrared cameras are labor‑intensive, inconsistent, and prone to human error. In response, researchers at Shanghai University of Electric Power have developed an advanced image segmentation method that combines an improved two‑dimensional Otsu thresholding technique with an enhanced region‑growing algorithm. This approach enables autonomous, real‑time defect identification in infrared images captured by inspection robots, significantly improving reliability and reducing false alarms.
The core challenge in pipeline defect detection lies in the complex visual environment of power plant interiors. Infrared images of high‑temperature pipelines often suffer from low contrast, cluttered backgrounds, and overlapping structures, making it difficult for conventional algorithms to distinguish between normal thermal patterns and genuine anomalies. Moreover, multiple defects may appear simultaneously in a single frame, requiring a segmentation strategy that can handle multi‑region analysis without manual intervention.
To address these issues, Peng Daogang, Yin Lei, Qi Erjiang, and their collaborators propose a two‑stage pipeline that first isolates the pipe structure from the background and then precisely extracts defective zones. The initial stage employs a modified two‑dimensional Otsu method—known for its ability to consider both pixel intensity and local neighborhood information—to perform pre‑segmentation. Unlike the classic one‑dimensional Otsu, which relies solely on global intensity histograms and often fails under noisy conditions, the two‑dimensional variant constructs a joint histogram of pixel gray values and their 3×3 neighborhood averages. This captures spatial context and enhances the separation between foreground and background.
The team further refined this approach by incorporating gray‑level magnitude into the joint probability density function and introducing an intra‑class variance term into the threshold optimization objective. The resulting criterion maximizes inter‑class variance while minimizing intra‑class dispersion, yielding a more robust and noise‑resilient threshold. In practice, this step effectively suppresses irrelevant background elements—such as support structures, cables, and ambient heat sources—leaving only the pipe region with preserved intensity values.
Once the pipeline is isolated, the second stage focuses on defect localization and segmentation. The algorithm begins by analyzing the gray‑level histogram of the pre‑segmented image to determine the dominant intensity value, denoted as T, which corresponds to the normal operating temperature of the pipe. A tolerance margin, defined as e = 1.4 × T, accounts for natural thermal fluctuations. Pixels whose 5×5 neighborhood average exceeds T* + e are flagged as potential defect centers.
This automatic seed‑point selection mechanism eliminates the need for human input and adapts dynamically to varying defect intensities across different scenes. After identifying a candidate seed, the system initiates a region‑growing process governed by two complementary criteria: an adaptive intensity threshold and a gradient‑based edge constraint.
The adaptive threshold is derived from the current mean and standard deviation of the growing region’s pixel intensities. Specifically, as the region expands, its statistical properties are continuously updated, and the allowable deviation for new pixels is adjusted accordingly. This self‑tuning behavior prevents both over‑segmentation (where healthy areas are mistakenly included) and under‑segmentation (where parts of a defect are excluded due to intensity drop‑offs near edges).
To further refine boundary accuracy, the researchers integrated a Prewitt operator–based gradient constraint. Since defect boundaries typically exhibit sharp thermal transitions, the magnitude of the local gradient serves as a reliable indicator of edge presence. The algorithm computes the gradient amplitude at both the current seed and each candidate neighbor, allowing inclusion only if the difference remains below a predefined threshold (empirically set between 30 and 50 for 8‑bit images). This dual‑criterion approach ensures that the segmented region remains both thermally coherent and geometrically well‑defined.
The team validated their method using a dataset of infrared images captured in real power plant environments under varying operational conditions. Three representative scenes—featuring one, two, and four distinct defects—were selected to test robustness against background complexity, edge ambiguity, and multi‑defect coexistence. The pipeline was implemented in C++ using OpenCV 4.1.0 and executed on a standard workstation (Intel Core i5‑4200M, 4 GB RAM, Windows 7 64‑bit).
Results demonstrated that the proposed method accurately detected all defects without false positives. In Scene One, a single hot spot was segmented with 929 pixels—nearly identical to manual annotation. Scene Two, containing two adjacent anomalies, yielded 796 and 489 pixels respectively, closely matching ground truth dimensions. Even in the most challenging scenario—Scene Three, with four small, closely spaced defects amidst a dense pipe network—the algorithm successfully isolated each region (249, 129, 109, and 96 pixels) without merging or fragmentation.
Comparative experiments against both the traditional region‑growing algorithm (with fixed thresholds) and a recent automatic method from the literature revealed clear advantages. The fixed‑threshold approach suffered from under‑segmentation in low‑contrast regions and over‑segmentation near blurred edges, while the literature method, though adaptive, produced false detections in noisy areas. In contrast, the new algorithm maintained high precision across all test cases.
Processing time averaged 0.331 seconds per image, well within the real‑time requirements of robotic inspection systems that typically operate at 1–2 frames per second. Although slightly slower than the fastest baseline (0.254 s), the trade‑off in accuracy and autonomy is justified for safety‑critical applications. Notably, the entire workflow—from image ingestion to defect segmentation—requires no human intervention, a crucial feature for large‑scale deployment in unmanned power plants.
Beyond technical performance, the methodology aligns with emerging industry trends toward intelligent operations and predictive maintenance. As utilities invest in digital twins and AI‑driven asset management, reliable defect detection becomes a foundational capability. Integrating this algorithm into robotic inspection platforms could enable continuous monitoring, early warning, and automated work order generation, thereby reducing unplanned outages and extending asset life.
The research also opens avenues for future enhancements. While the current implementation focuses on grayscale infrared data, future work could incorporate multi‑spectral or time‑series thermal imaging to improve discrimination between transient thermal spikes and persistent defects. Additionally, coupling the segmentation output with machine learning classifiers might enable root‑cause analysis—distinguishing between insulation failure, flow obstruction, or mechanical stress based on thermal morphology.
Hardware integration remains the next logical step. Although the algorithm has been validated offline, embedding it into an embedded processor on a mobile inspection robot would test its performance under real‑time constraints, communication latency, and environmental variability (e.g., vibration, dust, and electromagnetic interference). Field trials in operational power stations would provide valuable feedback on long‑term robustness and maintenance impact.
From a broader perspective, this work exemplifies the synergy between classical computer vision techniques and domain‑specific engineering challenges. Rather than relying solely on deep learning—which often demands large annotated datasets and significant computational resources—the team leveraged well‑understood principles (Otsu thresholding, region growing) and augmented them with targeted innovations tailored to the physics of thermal imaging and pipeline behavior. This hybrid strategy offers a pragmatic path toward trustworthy, interpretable, and deployable AI in industrial settings.
In an era where energy infrastructure faces increasing pressure to operate safely, efficiently, and sustainably, such advances in autonomous inspection represent more than just algorithmic progress—they are enablers of resilient, next‑generation power systems. As smart plants evolve from concept to reality, vision‑based defect detection will play an indispensable role in safeguarding critical thermal assets.
Peng Daogang, Yin Lei, Qi Erjiang, Hu Jie, Yang Xiaowei. Power Plant Pipeline Defect Detection and Segmentation Based on Otsu’s and Region Growing Algorithms. Infrared Technology, Vol. 43, No. 5, May 2021, pp. 502–509. DOI: 10.11846/IR2021430500502