A no-code artificial-intelligence tool for crack detection produced a mixed result in a small test of tungsten microscopy images. YOLOEZ recorded higher recall, F1 score and intersection over union than a tuned classical image-processing pipeline, indicating that it found more of the reference cracks and matched their marked areas more closely. The classical baseline, however, had higher precision and specificity, so YOLOEZ's advantage did not extend to every measure.
Inside the test
The study paired a qualitative software-feature comparison with a quantitative benchmark. The feature review described a tradeoff between the accessibility and flexibility of the evaluated tools. The benchmark targeted detection and localization of micro-scale cracks in scanning electron microscope, or SEM, images of additively manufactured tungsten.
The image set was divided into 20 training, 5 validation and 15 test images. Each of the 25 training and validation images was augmented four times, producing 100 additional synthetic samples. Ten test images came from a separate tungsten sample.
A pretrained medium-sized YOLO11 segmentation model was fine-tuned to convergence against manual crack annotations. Ten independent models supplied 95% confidence intervals for the YOLOEZ metrics.
On the 15 test images, YOLOEZ's recall was 0.6065, compared with 0.4372 for the baseline. Recall captures how many defects in the reference annotations were found, so the higher value indicates that YOLOEZ identified more of those cracks. Its precision was 0.5820, below the baseline's 0.7403, meaning a larger share of its detections did not count as correct. YOLOEZ's reported intervals were 95% confidence intervals of plus or minus 0.032 for recall and plus or minus 0.044 for precision, calculated from 10 independent training runs.
YOLOEZ also had a higher F1 score, a combined measure of recall and precision, at 0.5591 plus or minus 0.027 versus 0.5353 for the baseline. Its IoU, which measures how closely predicted crack areas overlap the manual markings, was 0.4150 plus or minus 0.024 versus 0.3962. Precision and specificity favored the baseline.
Crack counts showed the same tension. YOLOEZ detected 84 cracks, with a reported 95% confidence interval of plus or minus 10, while the image-processing baseline detected 129; the ground-truth count was 63. YOLOEZ's count was closer to the reference, even though it did not match it exactly. Both methods were assessed with pixel-level and object-level measures, and crack-free test images were excluded from average recall, precision, F1 and IoU calculations.
Why the comparison was not a clean win
The researchers tuned the morphological baseline on the training and validation images before testing it. They evaluated 780 parameter combinations using F1 score and IoU, giving the classical method a selected setting for the comparison.
That process also showed how sensitive the baseline was to its settings. Across the parameter search, training-set F1 ranged from 0.26 to 0.65 and recall from 0.17 to 0.58. The Otsu scaling factor produced mean spreads of 0.33 in F1 and 0.35 in recall. These were observed search ranges, not confidence intervals.
YOLOEZ was introduced as an open-source graphical interface intended to bring image labeling, YOLO model training and inference into one code-free workflow for structural-health-monitoring defect detection. The authors presented it as an integrated option within the accessibility-flexibility tradeoff identified in the software comparison.
A broader test still lies ahead
YOLOEZ was trained for 250 epochs. Over that period, training and validation segmentation loss generally fell, while validation precision, recall and mAP50(M) generally rose. The validation set contained only five images, so individual images could substantially change the epoch-level aggregate metrics.
The separate tungsten sample provided 10 out-of-distribution test images, meaning images outside the sample used for training. The authors interpret representative detection output as suggesting that the model learned features that generalize beyond the specific training sample. No aggregate out-of-distribution effect estimate was reported, leaving this as a qualitative indication rather than a broad performance measure.
The study also does not settle whether the interface is easier for people to use. Its feature comparison relied on software documentation rather than a direct user study, and no formal usability analysis was reported. Formal between-method significance tests were also absent, while baseline confidence intervals were not reported.
What is available now
The document is labeled a preprint, although it states that it is the authors' version of a paper accepted for publication in the ASME 2026 conference proceedings. The paper says YOLOEZ is available as an open-source project. It also says in-distribution images will be released through the Purdue University Research Repository, while out-of-distribution images will become available with a companion publication.
The work acknowledges support from the National Science Foundation and the U.S. Department of Energy, including its Office of Science and DOE Fusion Energy Science programs. Within the reported evidence, the central result is an image benchmark with a measurable split: YOLOEZ had higher recall, F1 score and IoU, while the baseline had higher precision and specificity.
Paper data and sources
Original title: Lowering the Barrier to AI-Driven Inspection: A No-Code Workflow for Automated Structural Defect Detection
Authors: Michael Holm, Tanner McElroy, Xinghang Zhang, Guang Lin
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text