Source-linked AI summary
Lowering the Barrier to AI-Driven Inspection: A No-Code Workflow for Automated Structural Defect Detection
Michael Holm, Tanner McElroy, Xinghang Zhang, Guang Lin
TL;DR
Structural health monitoring needs scalable defect detection, but fragmented workflows and technical requirements limit accessibility for domain experts. This paper introduces YOLOEZ, an integrated no-code YOLO platform that outperforms a classical morphological baseline across several detection metrics.
Problem
SHM defect detection remains difficult to scale because existing workflows require specialized expertise and integrating multiple components can hinder reproducibility.
Method
YOLOEZ integrates data labeling, YOLO model training, and inference in a single GUI-based, code-free workflow for SHM applications.
Results
YOLOEZ outperforms a classical morphological baseline across recall, F1 score, IoU, and crack count accuracy despite training on only 20 labeled images.
Takeaways & Limitations
YOLOEZ lowers the barrier for domain experts to develop and deploy inspection models in digital twin and predictive maintenance workflows.
Takeaways & Limitations
YOLOEZ currently lacks multi-class support, limiting applications requiring simultaneous detection of multiple defect types.
Abstract
from arXiv · showhide
Structural health monitoring (SHM) is essential in modern engineering, providing data for condition-based maintenance, lifecycle assessment, and predictive decision-making. Traditionally, SHM relied on visual inspection to detect defects such as cracks and deformations. Early computer vision (CV) methods, including thresholding, edge detection, and handcrafted features, aimed to automate this process but were highly sensitive to noise, imaging variations, and multiscale defects, limiting their reliability. Recent advances in machine learning, particularly convolutional neural networks (CNNs) and You Only Look Once (YOLO), have improved defect detection accuracy and enabled real-time analysis. However, adoption in SHM remains limited due to technical barriers such as data labeling, model training, and deployment, which typically require programming expertise. To address this gap, we introduce YOLOEZ, an open-source, GUI-based tool for end-to-end YOLO model application. YOLOEZ integrates data labeling, training, and inference into a single interface, enabling high-performance model development without code while supporting reproducible workflows. Evaluation against existing software and classical image processing demonstrates that YOLOEZ not only outperforms traditional methods across most detection metrics, but also lowers adoption barriers present in other modern CV tools. By combining accuracy with accessibility, YOLOEZ facilitates wider use of AI-driven monitoring for predictive maintenance, digital twins, and intelligent structural systems.
1. INTRODUCTION
Structural health monitoring needs reliable, scalable inspection methods, but conventional visual inspection and early computer-vision approaches face practical and technical limitations. YOLOEZ addresses the accessibility gap by providing an open-source GUI for integrated, reproducible YOLO-based structural defect detection workflows.
- Structural health monitoring supports condition-based maintenance, lifecycle assessment, and predictive decision-making, while increasingly complex structures require reliable, scalable integrity assessment.
- Classical image-processing methods such as thresholding, edge detection, and handcrafted features are sensitive to noise, imaging conditions, and multiscale defects.
- Deploying modern computer-vision systems involves data labeling, model configuration, and inference-pipeline development, often requiring specialized technical expertise.
- YOLOEZ is an open-source GUI tool that unifies data labeling, YOLO model training, and inference for structural defect detection.
- YOLOEZ promotes reproducible experimentation through a consistent framework for dataset generation, model development, and inference, with open-source release and documentation.
2. BACKGROUND
Computer vision for SHM has progressed from interpretable but fragile classical image processing to deep-learning detectors such as CNNs and YOLO. However, fragmented, technically demanding workflows continue to limit adoption and reproducibility, motivating integrated GUI-based systems.
- Classical SHM defect-detection methods use thresholding, edge detection, region growing, and morphological operations because they are inexpensive, interpretable, and easy to implement.
- Complex defect geometries, varying scales, low contrast, heterogeneous backgrounds, and imaging variation degrade classical-method performance, often requiring extensive manual tuning.
- CNNs learn hierarchical features for complex visual tasks, while YOLO performs single-stage bounding-box and class-probability prediction in one forward pass for real-time inference.
- Developing YOLO models requires coordinated data collection, annotation, formatting, configuration, training, validation, and deployment, creating barriers for engineering experts without machine-learning training.
- Ultralytics, Roboflow, and annotation platforms provide useful capabilities but leave users managing programming, external training and inference, or multiple components, which can undermine reproducibility.
- GUI-based systems can hide implementation details, standardize workflows, improve accessibility and reproducibility, and reduce errors, although existing solutions remain incomplete.
3. METHODOLOGY
YOLOEZ was developed as a code-free, integrated platform for labeling, training, and inference, and evaluated through workflow comparisons and quantitative micro-crack detection experiments. The technical evaluation used SEM images of additively manufactured tungsten, compared YOLOEZ with a tuned morphological baseline, and assessed pixel- and object-level metrics.
- YOLOEZ platform: YOLOEZ integrates data labeling, YOLO model training, and inference in one graphical interface without requiring code-based interaction.The interface supports an end-to-end workflow from raw images through labeling, configuration, training, and inference.
- Evaluation design: The study evaluates YOLOEZ qualitatively for workflow efficiency, usability, and accessibility, and quantitatively for detecting and localizing micro-scale cracks in SEM images.The workflow comparison included Ultralytics, PyTorch-based YOLO implementations, and Roboflow, using GUI, labeling, training, inference, and integration characteristics.
- Technical evaluation: A medium-sized YOLO11 segmentation model was fine-tuned until convergence on SEM images of micro-cracked, additively manufactured tungsten components.The specimens were 8 mm × 8 mm × 5 mm pure tungsten blocks produced by laser powder bed fusion and imaged after cross-sectioning and polishing.
- Dataset and augmentation: The dataset comprised 20 training, 5 validation, and 15 test images, with the 25 training and validation images each augmented four times to produce 100 additional synthetic samples.The images included heterogeneous texture, near-resolution-limit crack widths, and variable crack morphology and density.
- Classical baseline: The classical baseline used grayscale normalization, Gaussian blurring, Otsu thresholding, morphological opening and closing, and contour filtering by area and aspect ratio.The pipeline used a 5 × 5 Gaussian kernel, threshold scaling factor 1.35, a 3 × 3 circular structuring element, area limits of 91 pixels to 10% of image area, and aspect ratio ≥1.5.
- Performance metrics: Both methods were assessed using recall, precision, F1 score, specificity, intersection over union (IoU), and total crack count.The morphological pipeline was tuned across 780 parameter combinations using training and validation F1 score and IoU before testing on the test subset.
4. RESULTS
The results compare YOLOEZ with existing YOLO tools and a morphological crack-detection baseline. They also characterize YOLOEZ training behavior over 250 epochs and select the best validation checkpoint for test evaluation.
- Tool comparison: Table 1 qualitatively compares YOLOEZ, Ultralytics, Roboflow, and PyTorch for data labeling, model training, and inference.The comparison also considers integrated GUI support for dataset management, training configuration, and result visualization.
- Detection performance: Table 2 reports pixel-level and object-level performance for YOLOEZ and a morphological baseline across 15 test images.Images without cracks were excluded from average recall, precision, F1 score, and IoU calculations because those metrics are undefined or uninformative without ground-truth cracks.
- Training behavior: YOLOEZ training and validation behavior is shown over 250 epochs, including loss, validation precision, recall, and mAP50(M).The figure describes a representative run among ten model training runs.
- Checkpoint selection: The best checkpoint was selected by the highest recorded validation mAP50(M) and used for all reported test evaluations.The reported curves indicate a relatively small training–validation loss gap and minimal overfitting despite only 20 labeled training images.
5. DISCUSSION
YOLOEZ occupies a distinct niche by combining accessibility, local and code-free workflows, while outperforming the morphological baseline on most detection metrics. Its remaining limitation is single-class support, but its accessibility supports digital-twin and predictive-maintenance workflows.
- Accessibility and flexibility: YOLOEZ uniquely combines a graphical interface, built-in labeling, local data management, and code-free operation in one tool.This combination targets SHM researchers with materials-characterization expertise but limited machine-learning experience.
- Accessibility and flexibility: Ultralytics and PyTorch provide greater architectural and training flexibility but require programming knowledge, while Roboflow raises cloud-governance concerns and reserves training for paid users.Ultralytics HUB offers a graphical interface, but advanced features and cloud-training quotas are restricted.
- Generalization and limitations: YOLOEZ currently supports only single-class crack detection, limiting future applications requiring simultaneous detection of cracks, voids, and delaminations.This limitation has no practical impact on the evaluated single-class task and identifies multi-class support as a future development direction.
- Quantitative comparison: YOLOEZ outperformed the morphological baseline across all metrics except precision and specificity despite training on only 20 labeled images.95% confidence intervals from ten independent training runs indicate that the advantage was consistent across initializations.
- Quantitative comparison: 0.6065 ± 0.032 vs. 0.4372: YOLOEZ achieved higher recall than the morphological baseline, identifying more true crack pixels and offering greater robustness to imaging and crack variations.The narrow recall confidence interval suggests this robustness was reliable across training runs.
- Quantitative comparison: The morphological baseline achieved higher precision, 0.7403 vs. 0.5820 ± 0.044, whereas YOLOEZ achieved high specificity, 0.9934 ± 0.002, statistically comparable to 0.9951.The baseline’s conservative behavior improved precision but reduced recall; both methods avoided widespread background false positives.
- Quantitative comparison: YOLOEZ detected 84 ± 10 objects versus 129 for the morphological method and achieved IoU of 0.4150 ± 0.024 versus 0.3962, indicating better instance and spatial correspondence.The morphological method over-segmented cracks and flagged surface-texture artifacts, while YOLOEZ more closely matched the ground-truth count of 63.
- Generalization and limitations: YOLOEZ maintained meaningful out-of-distribution detection performance with tight confidence intervals, while training curves indicated no significant overfitting despite the small dataset.Ten of 15 test images came from an unseen 3D-printed tungsten sample; validation-curve variance reflected only five validation images.
6. CONCLUSION
YOLOEZ is an open-source, GUI-based platform for end-to-end YOLO model development that lowers barriers to AI-driven structural defect detection for domain experts without programming expertise. It addresses a gap in SHM tools by combining graphical operation, built-in data labeling, local data management, and code-free workflow in one platform.
- Contribution: YOLOEZ provides end-to-end YOLO model development through an open-source graphical interface for domain experts without programming expertise.The platform is designed to lower the barrier to AI-driven structural defect detection.
- Contribution: YOLOEZ is the only compared platform combining a graphical interface, built-in data labeling, local data management, and code-free operation in a single SHM workflow.The qualitative comparison identifies this combination as a gap in the current ecosystem for structural health monitoring.