Abstract
Concrete surface crack inspection is increasingly supported by lightweight deep models, yet full automation remains difficult under field conditions where cracks are slender, low-contrast, and confounded by stains, shadows, rough textures, and motion artifacts. This article develops a human-in-the-loop framework for concrete crack inspection that combines lightweight vision models with calibrated uncertainty routing. The evidence base is deliberately narrow: bridge and concrete crack detection or segmentation studies using lightweight detectors and segmentation networks under complex backgrounds, a review of remote bridge inspection workflows, and inductive conformal prediction for class-label uncertainty in object detection. The article is a targeted narrative synthesis rather than a systematic review. It asks how automated crack screening can be organized so that machine efficiency is retained while ambiguous cases are explicitly routed to human review. The proposed method uses a two-stage vision pipeline: lightweight detection for candidate localization and lightweight segmentation for pixel-level delineation, followed by conformal prediction to quantify class-label uncertainty and operational triage rules that assign each candidate to automatic acceptance, automatic rejection, or expert review. Exact operational definitions are provided for crack candidates, uncertainty sets, triage states, and review triggers. A prospective validation protocol is specified, including calibration design, comparator conditions, outcome measures, subgroup analyses, and failure analysis. Conceptually, the manuscript establishes that uncertainty-calibrated routing is a plausible way to integrate lightweight crack models into remote concrete inspection workflows, but its field performance, calibration stability, and impact on inspector workload remain untested.
1. Introduction
Concrete surface cracks are common visible indicators of structural deterioration and are therefore central targets in bridge inspection. The cited bridge-inspection and crack-detection literature supports two premises relevant to this article. Manual inspection remains important, yet it is labor intensive, subjective, and sensitive to accessibility and lighting. At the same time, vision-based automation has advanced through lightweight object detectors and segmentation networks designed for crack screening under complex concrete backgrounds. [1] [2] [3]
The remaining problem is not only how to detect cracks, but how to decide when an automated output should be trusted without review and when it should be deferred to a human inspector. The crack-detection studies considered here show that modern lightweight models can improve localization and segmentation under difficult backgrounds, but they also state limits in dataset coverage, field-condition variability, and the difficulty of fine cracks, irregular morphologies, shadows, stains, and low contrast. Those limits matter operationally because an inspection workflow must govern ambiguous cases, not merely report average model accuracy. [2] [4] [5]
This article develops a bounded contribution to that operational problem. It proposes a human-in-the-loop protocol for concrete surface crack inspection in which lightweight computer vision performs first-pass screening and calibrated uncertainty determines whether a case is auto-accepted, auto-rejected, or routed to human review. The framing is restricted to visible concrete surface cracks in image-based inspection and excludes non-vision fault diagnosis, non-concrete defects, and claims of completed empirical validation. [1] [6]
2. Research question and evidence scope
The research question is: how can a concrete crack inspection pipeline combine lightweight detection and segmentation with explicit uncertainty calibration so that difficult cases are systematically referred to human inspectors rather than silently over-trusted? This question joins two strands of evidence that are usually discussed separately: crack-specific model design for complex backgrounds and uncertainty quantification for object detection. [2] [3] [6]
The evidence scope is a targeted narrative synthesis rather than a systematic review. Direct application evidence is drawn from four studies on bridge or concrete crack detection and segmentation: a lightweight semantic segmentation model coupled with detection under UAV inspection, two improved YOLO-family crack detectors designed for complex concrete backgrounds, and a lightweight collaborative detector for resource-constrained scenarios. Methodological foundations are drawn from a review of remote bridge inspection workflows and from an inductive conformal-prediction study that establishes finite-sample class-label coverage for the semantic classification component of object detection after detector score adaptation. That conformal result does not by itself validate localization, duplicate suppression, or the full crack-inspection workflow proposed here. [3] [2] [5] [4] [1] [6]
The article does not argue that object-detection conformal methods have already been validated for concrete crack inspection. Instead, it transfers the formal mechanism of calibrated class-label uncertainty to a crack-screening workflow and specifies how such a transfer should be tested prospectively. [6]
3. Prior evidence relevant to a human-in-the-loop crack workflow
The direct crack-detection studies support the need for specialized visual models rather than generic thresholding or edge operators alone. In complex concrete scenes, cracks may be weak, discontinuous, irregular, and easily confused with rough texture, stains, water traces, and shadows. This motivates architectural changes that preserve weak crack signals, strengthen edge representation, and improve multi-scale feature interaction. YOLO11-FR addresses full-spectrum feature interaction and edge enhancement for crack screening with bounding boxes, explicitly targeting fragmented localization and false activation from textured backgrounds. YOLOv8-CA-EYHL combines filtering-equalization preprocessing, coordinate attention, and enhanced feature learning to improve detection across orientations, scales, and crack morphologies. MCL-YOLO addresses slender and weak-texture cracks while reducing computation, with explicit concern for edge deployment. Feather SFASwinUnet focuses on semantic segmentation and fine-grained crack recovery, then combines segmentation with a detector for complex backgrounds and quantitative measurement from segmented masks. [2] [5] [4] [3]
These studies also imply an important division of labor between detection and segmentation. Object detectors are practical for rapid screening because they localize candidate regions with relatively low computational burden. Segmentation networks provide pixel-level masks needed for width, length, and area estimation, but they generally require denser annotation and may be costlier to deploy. The reviewed crack studies align with this distinction: some prioritize rapid screening with bounding boxes, while the segmentation study extends toward geometric measurement after localization. [2] [5] [3]
The remote inspection review places such models in a larger workflow. It distinguishes remote visual inspection from SHM, treats computer vision as one element in bridge digitalization, and notes that methods differ in maturity and field integration. It also reports a streamlined inspection system in which human-in-the-loop machine learning was used to segment structural elements before damage identification and integration into a georeferenced bridge information model. This is relevant because a human-in-the-loop crack detector should not be conceived as an isolated classifier. It is part of a data-production chain that must preserve image context, location, and traceable review decisions. [1]
Uncertainty handling is the missing operational layer. The conformal-prediction study is not about cracks, but it directly addresses class-label uncertainty in object detection. By adapting detectors to output per-class score vectors and a background-confidence estimate, it enables inductive conformal prediction with finite-sample coverage guarantees for class labels under standard assumptions. A practical implication for the present article is narrower than end-to-end detection validity: prediction-set size within that conformal framework can be used as an uncertainty signal, and larger or empty sets can be routed to targeted human review. That principle is transferable to crack screening, provided that candidate generation and localization are specified separately. [6]
The evidence is not uniform in its implications. Some crack studies report strong robustness or generalization within their evaluation settings, including cross-domain performance from laboratory to UAV pavement imagery or external validation on a second dataset. At the same time, the same literature acknowledges residual concerns about bridge diversity, acquisition devices, adverse weather, low illumination, wet surfaces, motion blur, and the difference between bridge-specific and non-bridge datasets. The appropriate synthesis is therefore not that current models are unreliable, but that they remain conditionally reliable and require workflow safeguards when transferred across inspection settings. [3] [2] [5]
4. Conceptual framework and operational definitions
The proposed framework treats crack inspection as a selective automation problem. The automated system should act autonomously only when its prediction is both sufficiently confident and sufficiently informative for the intended inspection task. Otherwise, it should preserve the candidate region and defer judgment to a human inspector. This formulation differs from maximizing detector accuracy alone because the operational target is a three-way partition of cases: accepted crack findings, rejected non-crack findings, and review-required findings. [6] [1]
A crack candidate is defined as any image region localized by the detector with box b and detector score vector s = (s1, ..., sK), where K is the number of semantic classes available to the detector. For the narrow use case here, the minimal class design is K = 2 with classes {crack, background-like non-crack}; a richer design may divide cracks by morphology, but that is optional and should not be assumed without data support. Each accepted crack candidate may be passed to a segmentation model that outputs a binary mask M over the cropped region. A mask-derived measurement record may then include length, width, and area estimates if the segmentation quality is adequate for measurement. This article does not assume that every detected crack qualifies for measurement. [3] [6]
Inductive conformal prediction is applied after candidate generation rather than to raw images. Let the detector produce an initial set of boxes with per-class score vectors and a background-confidence estimate. Duplicate and highly overlapping proposals are then resolved by a fixed suppression rule before conformalization: within each image, sort proposals by crack-class score and apply non-maximum suppression with one pre-registered IoU threshold, denoted tau_NMS; by default tau_NMS = 0.50, with sensitivity analysis at 0.30 and 0.70. The surviving boxes are the crack candidates. For each candidate x, inductive conformal prediction produces a prediction set C(x) over the detector classes. The primary triage rule is exact: route x to human review if |C(x)| != 1. A singleton set {crack} is eligible for automatic acceptance, a singleton set {non-crack} is eligible for automatic rejection, and any multi-label or empty set triggers review. This rule uses conformal uncertainty only for class-label triage; it does not claim conformal validity for localization quality. [6]
Segmentation checks are applied only to candidates with C(x) = {crack}. Let M be the binary mask predicted on the cropped candidate region after resizing that crop to a fixed reference resolution before segmentation, so that derived shape measures are comparable across devices. Define A(M) as mask area in pixels at the reference resolution and L(M) as one-pixel skeleton length. Use the scale-normalized thinness ratio Q(M) = L(M)^2 / max(A(M),1), which is dimensionless and therefore preferable to L(M)/A(M) for cross-platform evaluation. On the calibration split, estimate the admissible interval [q_low, q_high] from accepted crack masks by the empirical alpha_shape/2 and 1-alpha_shape/2 quantiles, with default alpha_shape = 0.05 and sensitivity analysis at 0.10. If C(x) = {crack} but Q(M) falls outside this interval, or if the segmenter fails to return a stable mask after one repeat inference on the same crop, the candidate is routed to review rather than auto-accepted. No mask-based rule is used to auto-reject a candidate already labeled {non-crack}. [3] [6]
The triage state is therefore hierarchical. First apply conformal class-label triage to the non-duplicated candidate set. If C(x) = {non-crack}, set T(x) = reject. If |C(x)| != 1, set T(x) = review. Only if C(x) = {crack} does the protocol evaluate auxiliary checks. When segmentation is inactive, T(x) = accept for C(x) = {crack}. When segmentation is active, T(x) = accept only if C(x) = {crack} and the mask-quality checks pass; otherwise T(x) = review. This ordering prevents conflicts between rejection and mask plausibility, and it keeps every auxiliary rule subordinate to the primary uncertainty gate. [6] [3]
5. Proposed methodology
The proposed methodology has six stages: image acquisition, candidate localization, optional candidate segmentation, conformal calibration, triage decision, and review integration. Image acquisition may use handheld, fixed, or UAV-mounted RGB cameras, but the method is limited to visible-surface inspection. The remote inspection review supports the practical use of camera-based methods and their integration with broader digital workflows, while also distinguishing these visual methods from SHM sensors that capture structural response rather than visible surface damage. [1]
Stage 1, candidate localization: a lightweight object detector identifies regions that may contain cracks and outputs per-class scores for each proposal. YOLO-family detectors are reasonable starting points because the cited crack literature repeatedly motivates them for rapid screening and deployment efficiency. The protocol is detector-agnostic, but comparisons should be framed as selected design instantiations rather than as a field standard. A prospective study may compare a lightweight baseline detector with crack-oriented variants such as YOLO11-FR, YOLOv8-CA-EYHL, or MCL-YOLO, because these models were each designed to address slender cracks, weak boundaries, or complex backgrounds through different architectural mechanisms. [2] [4] [5]
Stage 2, candidate segmentation: each detector proposal classified as potentially crack-containing is cropped and passed to a lightweight segmentation model when pixel-level delineation is needed. Feather SFASwinUnet is the clearest source-grounded prototype because it was explicitly designed to recover fine-grained cracks with low overhead and was combined with detection under complex backgrounds. In the proposed workflow, segmentation is not mandatory for every inspection objective. If the operational goal is rapid screening only, segmentation may be skipped; if the goal includes quantitative crack measurement, segmentation becomes necessary. [3] [2]
Stage 3, conformal calibration: the detector must expose per-class scores and a background or non-crack confidence estimate, because conformal class-label coverage requires more than a top-1 score. The cited conformal study demonstrates score adaptation for several detector families, including YOLO11, but it does not validate the crack-specific routing policy proposed here. Calibration is performed on a split disjoint from training and final evaluation, after candidate generation and duplicate suppression have been fixed. For each retained candidate matched to a reference label under a pre-registered assignment rule, compute nonconformity scores and fit the conformal threshold. APS is the default nonconformity function because the cited study reports favorable coverage-efficiency trade-offs relative to several alternatives; Hinge is the prespecified sensitivity alternative. Because calibration validity may degrade under score-distribution shift, recalibration is mandatory after detector changes and should also be repeated for each new site-camera configuration before autonomous use. [6]
Stage 4, triage decision: the conformal module outputs a prediction set at target coverage 1 - alpha. The default operating point is alpha = 0.10. Sensitivity analysis may additionally evaluate alpha = 0.05 without altering the decision logic. The triage rule is fixed: singleton {crack} candidates proceed to acceptance checks, singleton {non-crack} candidates are rejected, and all other sets are reviewed. When segmentation is active, only singleton {crack} candidates are segmented. Review is then triggered if the mask thinness ratio falls outside the calibrated interval, if detector and segmenter cannot be aligned to the same crop, or if stable mask extraction fails on one repeated inference. This structure isolates the effect of conformal triage from downstream mask-quality screening. [6] [3]
Stage 5, human review: reviewed candidates are presented with the original image, localized box, class scores, prediction set, and if available the segmentation mask and mask-derived measurements. Human review should support at least three actions: confirm crack, dismiss as non-crack, or mark indeterminate and request reacquisition. This is where the workflow differs from a conventional detector benchmark. The human is not a fallback after unrecorded model failure; the human is a planned decision node with documented reasons for override or confirmation. [1]
Stage 6, structured integration: accepted and human-confirmed cracks should be linked to spatial and asset context so that remote inspection outputs can populate digital inspection records. The remote bridge-inspection review supports this emphasis by describing workflows that connect computer vision outputs to 3D models and bridge information models. The present proposal does not require full digital-twin infrastructure, but it assumes that image evidence, timestamps, camera viewpoint metadata, and review status are retained for traceability. [1]
6. Comparator conditions, variables, and outcome measures
A credible evaluation should compare the proposed pipeline against two precise comparators. Comparator A is detector-only automation: the same detector, candidate-generation rule, and duplicate-suppression rule are used, but decisions are made by a fixed detector threshold without conformal prediction and without mandatory human review of ambiguous candidates. Comparator B is detector-plus-segmentation automation: the same detector and the same segmentation model used in the proposed pipeline are retained, but triage is based on fixed detector and mask-quality thresholds rather than conformal prediction. This specification isolates the incremental effect of calibrated uncertainty routing from the separate effect of adding segmentation. [2] [3] [6]
Primary independent variables are detector architecture, segmentation on or off, nonconformity function, target coverage level, and review policy. Key controlled variables include imaging platform, annotation protocol, train-calibration-test split strategy, candidate-generation rule, and the duplicate-suppression threshold tau_NMS. Candidate-level evaluation is primary because triage decisions are made per candidate. Inspection-level summaries should then be derived by pre-registered aggregation rules: image-level false positives equal the count of auto-accepted or human-confirmed non-reference candidates per image after duplicate suppression, surface-level false positives equal the same count divided by inspected surface area when area metadata exist, and image status is positive if at least one candidate on that image is finally confirmed as a crack. These rules connect candidate triage outputs to image-level and inspection-level endpoints without changing the unit of decision. [6]
Primary outcomes are candidate-level crack sensitivity, candidate-level false acceptance rate among non-crack candidates, review rate, and final resolved accuracy after review. For inspection summarization, report false positives per image after duplicate suppression and, where surface area is known, false positives per inspected surface area. When segmentation is used, report segmentation usability only among candidates that reach the segmentation stage under the protocol. Human-process outcomes should include override rate for auto-accepted crack candidates and the proportion of reviewed candidates marked indeterminate and sent for reacquisition. Calibration outcomes should include empirical class-label set coverage, singleton-set frequency, multi-label-set frequency, and empty-set frequency. Prediction-set size is interpreted here as an operational workload signal within the conformal framework rather than as a universal property of detectors. [6]
For measurement-capable pipelines, secondary outcomes may include agreement between mask-derived crack dimensions and reference annotations, but only for candidates judged segmentable by the protocol. This restriction matters because the detector studies that use bounding boxes do not directly estimate crack width or length, whereas the segmentation study explicitly links masks to quantitative analysis. A fair evaluation should not penalize detector-only methods for tasks they were not designed to perform. [2] [3]
7. Prospective validation design
A prospective evaluation can replace a results section for the present conceptual article. The validation design should include at least four data partitions or collection strata: training, calibration, internal test, and external field test. The external field test is essential because the crack literature repeatedly warns that bridge types, sensors, lighting, wetness, shadows, motion blur, and texture conditions vary across sites and are not exhaustively covered by any single dataset. Laboratory-to-field or public-to-local transfer should therefore be treated as a distinct generalization question rather than assumed solved by high internal accuracy. [2] [3] [5]
Ground truth should be annotated at two linked levels: candidate existence and, where relevant, pixel-level extent. Candidate-reference matching must be pre-registered, for example by one-to-one assignment with an IoU threshold applied after duplicate suppression, so that conformalized candidates and evaluation labels are aligned. Review decisions by inspectors should be recorded independently before consensus adjudication so that disagreement can be analyzed explicitly. Because manual inspection is subjective, consensus is not treated as a perfect gold standard; it is a procedure for reducing single-rater variability. The remote-inspection review frames human-level reliability as an integration requirement for deployment into existing inspection systems, which implies that inter-inspector variability should be measured rather than ignored. [1]
Subgroup analysis should be pre-specified with operational labels rather than assigned informally after inspection. A candidate is classified as visually small and slender if the annotated crack mask, or the minimum bounding rectangle when only a box is available, has minor-axis width below a pre-registered pixel threshold at the reference resolution and aspect ratio above a pre-registered threshold. A candidate is classified as irregular or branched if the reference annotation contains at least one branch point on a one-pixel skeleton or a curvature measure above a pre-registered threshold. A candidate is classified as low-contrast if the mean absolute grayscale contrast between annotated crack pixels and a fixed-width surrounding band falls below a pre-registered threshold. Complex background is a contextual label applied when the surrounding region contains stains, shadows, rough texture, or water traces according to an annotation guide completed before outcome analysis. UAV imagery should be analyzed separately from close-range handheld imagery when both are used, because aerial acquisition introduces motion blur, perspective distortion, and uneven light. [4] [2] [3]
Ablation within the prospective protocol should answer four questions: whether conformal routing reduces unsafe auto-acceptance relative to fixed thresholding; whether segmentation adds useful measurement capability after uncertain detections or only adds review cost; whether the calibrated thinness-ratio check reduces texture-induced false positives among singleton {crack} candidates; and whether class-label coverage remains stable after platform changes. Calibration should therefore be repeated per detector and per site-camera configuration, not borrowed unchanged across devices or locations. Sample size for the calibration split should be fixed prospectively by a precision target for empirical coverage estimation, such as requiring the half-width of a binomial confidence interval around observed coverage to be below a pre-registered margin; the exact count then follows from the chosen target coverage and margin rather than from an ad hoc rule. [6]
8. Failure cases, assumptions, and limits of generalization
The proposal rests on assumptions that may fail in practice. Inductive conformal prediction relies on calibration data that are representative of deployment conditions. If the inspection environment changes sharply, empirical coverage may deteriorate even though the procedure is formally valid under its assumptions. This risk is especially salient for concrete inspection because shadows, moisture, corrosion products, and camera motion can alter the score distribution of weak crack targets. [6] [2]
The workflow also assumes that uncertainty in class label is a useful proxy for overall inspection uncertainty. That assumption is incomplete. A detector may be confident about the crack class yet still localize the wrong region, or a segmenter may produce an implausible mask from a correctly localized box. This is why the protocol supplements conformal class sets with detector–segmenter agreement and mask plausibility checks. Even so, these checks cannot fully resolve cases where multiple adjacent defects merge, where surface joints mimic cracks, or where non-crack linear markings have crack-like geometry. [4] [3]
Generalization is intentionally narrow. The proposal concerns visible concrete surface crack inspection, particularly in bridge-like settings with complex backgrounds. It should not be assumed to extend to subsurface deterioration, steel corrosion grading, asphalt-only defects, multimodal SHM diagnosis, or non-vision fault domains. The remote inspection review makes clear that bridge assessment encompasses many sensing and analysis modalities. The present article addresses one bounded task within that broader ecosystem. [1]
A further limit concerns workload economics. Human-in-the-loop systems are useful only if review burden remains manageable relative to safety gains. The cited literature supports routing uncertain cases to human inspection, but it does not establish an optimal review fraction for concrete cracks. That operating point will depend on crack prevalence, acquisition quality, staffing, and the tolerated rate of unsafe auto-acceptance. For that reason, the present article specifies one default operational rule with limited sensitivity analyses rather than a broad menu of thresholds, and every numerical setting remains hypothetical until prospectively evaluated. [6]
9. Conclusion
This article has developed a conceptual protocol for human-in-the-loop concrete crack inspection that integrates lightweight object detection, optional lightweight segmentation, and conformal uncertainty routing. The targeted synthesis supports three claims. First, lightweight detectors and segmentation networks are already being engineered to address crack-specific difficulties such as slender morphology, weak boundaries, and textured concrete backgrounds. Second, remote bridge inspection workflows increasingly require structured, traceable digital outputs rather than isolated model predictions. Third, inductive conformal prediction offers a principled way to convert detector scores into actionable uncertainty signals for routing ambiguous candidates to human review. [2] [5] [4] [3] [1] [6]
What has been established here is conceptual and methodological, not empirical. The manuscript provides operational definitions, exact triage rules, comparator conditions, outcome measures, and a prospective validation design. What remains untested is whether the proposed routing policy achieves acceptable coverage, preserves sensitivity to fine cracks, reduces unsafe automation, and maintains feasible reviewer workload across real concrete inspection settings. Those questions require prospective calibration and field evaluation before claims about effectiveness or deployment readiness can be justified. [6] [1]
Contribution and availability
The manuscript contributes a bounded conceptual framework for concrete surface crack inspection: a human-in-the-loop protocol that combines lightweight detection, optional segmentation for measurement, and inductive conformal prediction for explicit triage to automatic acceptance, automatic rejection, or human review. It specifies operational rules, variables, comparators, outcomes, and a prospective validation design tailored to crack inspection under complex concrete backgrounds.
Funding: Roman Vlasov’s own funds. Editorial responsibility and competing interests: Roman Vlasov is the founder and Editor-in-Chief of Smart World Journal and the creator of the BILDRY construction-planning platform. Other competing interests have not been declared.
This conceptual study reports no newly collected empirical data. Its proposed procedures require the prospective validation described above.
References
- Ribeiro, Diogo; Rakoczy, Anna M.; Cabral, Rafael; Hoskere, Vedhus; Narazaki, Yasutaka; Santos, Ricardo; Tondo, Gledson; Gonzalez, Luis; et al. (2025). Methodologies for Remote Bridge Inspection—Review. Sensors. 25(18), 5708. doi:10.3390/s25185708.
- Zhang, Yangming; Tian, Baohui; Guo, Hufeng (2026). YOLO11-FR: A bridge crack detection method based on frequency-domain fusion and an edge enhancement mechanism. PLOS One. 21(7), e0354254. doi:10.1371/journal.pone.0354254.
- Dong, Xuwei; Yu, Zhibing; Yuan, Jiashuo; Dai, Jinpeng (2026). Automatic bridge crack detections: A lightweight segmentation model for UAV inspection. PLOS One. 21(9), e0358865. doi:10.1371/journal.pone.0358865.
- Han, Bingyu; Wu, Yang; Feng, Wenhao; Mi, Xiaoman (2026). MCL-YOLO: A Multi-Module Collaborative Lightweight Object Detection Method for Bridge Crack Detection. Sensors. 26(18), 5801. doi:10.3390/s26185801.
- Zhu, Xianwei; Chao, He; Zhang, Yahui (2026). Yolov8-CA-EYHL: A bridge crack detection algorithm based on improved YOLOv8. PLOS One. 21(9), e0358872. doi:10.1371/journal.pone.0358872.
- Mohammed, Mohammed Aliy; Anzaku, Esla Timothy; Jonkers, Jef; Krishnamoorthy, Janarthanan; Neve, Wesley De; Van Hoecke, Sofie (2026). Inductive Conformal Prediction for Guaranteed Class-Label Coverage in Object Detection. Journal of Imaging. 12(8), 348. doi:10.3390/jimaging12080348.