Abstract
Bridge crack annotation is a bottleneck in image-based inspection because cracks are weakly contrasted, slender, irregular, and easily confounded by concrete texture, stains, shadow, and other surface patterns. Existing bridge-crack detectors and segmenters show that architectural changes can improve robustness, but they do not remove the cost of manual labeling. This protocol asks how active learning can reduce annotation burden while preserving detection and segmentation quality for bridge imagery. The proposed framework couples a bridge-inspection acquisition workflow with uncertainty- and diversity-driven sample selection, then routes selected images to bounding-box annotation, pixel-level crack masking, or both according to a prespecified task-value rule. The methodology defines an image-level informativeness score from normalized model uncertainty, morphology-sensitive redundancy, and bridge-context coverage, with explicit default weights and sensitivity ranges. The protocol is designed for detector-first and segmentation-first pipelines and includes comparators based on random sampling and uncertainty-only sampling, using the same annotation-routing logic across study arms. Outcome measures are annotation efficiency at matched annotation cost, detection and segmentation quality, calibration of uncertainty estimates, and cross-bridge generalization. Remote bridge-inspection methodology informs the acquisition and review workflow, and hierarchical Bayesian deterioration modelling provides an interpretive framework for underrepresented crack regimes and inspection uncertainty. The manuscript specifies assumptions, failure cases, stopping rules, and a prospective validation design, but it does not report empirical results.
1. Research question and evidence scope
This protocol addresses a specific operational question: how should an annotation system choose bridge-crack images for manual labeling so that subsequent detection and segmentation models improve with fewer labeled samples? The scope is limited to bridge imagery and to active-learning-oriented sample selection for crack detection and semantic segmentation. It excludes federated learning and non-bridge domains. The evidence base is a targeted narrative synthesis of prior bridge-crack detection and segmentation work, remote bridge-inspection methodology, and hierarchical Bayesian deterioration modelling. It is not a systematic review and does not claim exhaustive coverage or fixed search counts. [1] [2] [3] [4] [5]
The direct application evidence shows that bridge-crack detectors and segmenters already target the main visual difficulties relevant to annotation: weak contrast, slender or irregular morphology, and interference from complex textures. Detection studies have used filtering-equalization preprocessing, coordinate attention, high- and low-frequency feature learning, frequency-domain fusion, edge enhancement, and lightweight design choices to improve recognition of fine cracks under bridge-surface clutter. Segmentation work has combined global semantic capture with fine-detail recovery and has also integrated detection and segmentation to improve robustness under complex backgrounds. These studies motivate annotation policies that are sensitive not only to model uncertainty, but also to crack morphology and image complexity. [1] [2] [3] [6]
2. Prior work and the gap in annotation strategy
The bridge-crack literature demonstrates that the recognition problem is difficult, but it mostly addresses network design after labels already exist. YOLOv8-CA-EYHL combines filtering-equalization preprocessing with coordinate attention and enhanced high- and low-frequency feature learning to better represent weak, slender, and irregular cracks; YOLO11-FR adds frequency-domain interaction and residual edge enhancement to strengthen boundary cues; MCL-YOLO reduces computational load while using receptive-field attention and high-resolution semantic reinforcement to improve detection of microcracks. On the segmentation side, Feather SFASwinUnet introduces scale-fusion attention and a lightweight skip-connection block to preserve fine crack detail, and it combines detection and segmentation to improve robustness in complex backgrounds. [1] [3] [6] [2]
These methods support two conclusions relevant to annotation policy. First, bridge crack images vary in how much information they provide to a learner: some images contain obvious, repeated patterns, while others contain rare morphologies or difficult backgrounds. Second, low-contrast and slender cracks remain difficult even for improved architectures, which means annotation selection should not rely on easy-to-label images alone. The gap is therefore not only how to detect cracks, but which bridge images should be labeled first, and at which annotation granularity, so that learning data are spent on samples with the highest expected value. [1] [3] [6] [2]
3. Proposed framework and operating assumptions
The proposed framework has three components: acquisition, triage, and annotation. Acquisition follows remote bridge-inspection practice, in which camera-based inspection complements field access and supports image capture from bridge decks, piers, and abutments under varying viewpoints. Triage uses an active-learning policy to rank unlabeled images. Annotation assigns each selected image to one of three actions: bounding-box labeling only, segmentation-mask labeling only, or dual annotation when the image is expected to contribute to both tasks. This workflow is intended for bridge inspection teams that collect images iteratively over time, not for offline corpus construction only. [4] [1] [2]
The framework assumes that there is a seed set of labeled bridge images covering common crack appearances and that a base model can produce calibrated or approximately calibrated uncertainty estimates on unlabeled images. It also assumes that bridge images can be assigned metadata such as bridge component, viewing distance, illumination condition, surface condition, and acquisition source. These metadata do not need to be perfect, but they must be sufficiently stable to support diversity-aware sampling. If metadata are unavailable, the protocol still functions, but its diversity term collapses to feature-space clustering rather than condition-aware balancing. [4] [5]
A useful probabilistic interpretation comes from hierarchical Bayesian deterioration modelling. In that literature, inspection-derived deterioration risk and monitoring-derived crack evolution are treated as different layers of evidence. For this protocol, that analogy is interpretive rather than inferential: a candidate image is valuable not only when the current learner is uncertain, but also when the image represents a crack regime or inspection condition that is underrepresented in the labeled set. The protocol therefore treats each candidate image as proposed evidence about both visual recognizability and deterioration-relevant context, without claiming that image labels alone estimate structural state. [5]
4. Operational definitions and sample selection rule
Let U be the pool of unlabeled bridge images and L the current labeled set. For each image x in U, compute three normalized terms on the unit interval. First, define task-normalized uncertainty u(x)=0.5u_det(x)+0.5u_seg(x). To make u_det(x) reproducible, use test-time Monte Carlo uncertainty from T stochastic forward passes of the detector with dropout enabled in the detection head; the default is T=20. After non-maximum suppression within each pass, match detections across passes by an intersection-over-union threshold of at least 0.5 and retain the K highest-confidence matched detections in the aggregated result, with default K=10. For each retained detection q, let p_qt be its crack probability at pass t and define binary-entropy uncertainty h_q=H(mean_t p_qt)/log 2, where H(p)=-p log p-(1-p)log(1-p). Let u_det(x) be the mean of h_q over the retained detections, and set u_det(x)=0 when no detection is produced in any pass. Define u_seg(x) analogously as the mean normalized binary entropy over the pixelwise mean crack-probability map from T stochastic forward passes of the segmenter; if only one task model is available, assign the available term weight 1 and the missing term weight 0. Second, define redundancy r(x) as the mean cosine similarity between the embedding of x and its k=5 nearest neighbors in L, linearly rescaled to [0,1] from the cosine range [-1,1]. Third, define coverage deficit c(x)=1-(1/m)Σ_j p_Lj(z_j(x)), where z_j(x) is the category of x on metadata dimension j, p_Lj is the empirical relative frequency of that category in L, and m is the number of metadata dimensions used. The default acquisition score is s(x)=0.5u(x)+0.3c(x)-0.2r(x). Sensitivity analysis perturbs one weight at a time while preserving the sign pattern and unit sum. Specifically, for baseline weights w=(0.5,0.3,0.2) on uncertainty, coverage deficit, and redundancy magnitude, choose an index i and a perturbation δ in {-0.1,-0.05,0.05,0.1} such that 0≤w_i+δ≤1. Set w_i'=w_i+δ and rescale the remaining two weights proportionally to their baseline values: for j≠i, w_j'=(1-w_i')w_j/(1-w_i). The score under perturbation is then s'(x)=w_u'u(x)+w_c'c(x)-w_r'r(x). This procedure disallows negative weights and handles boundary cases by excluding any δ that would place a weight outside [0,1].
Selection proceeds by ranking all x in U by s(x) and then applying a duplicate filter. Let d(x,x')=1-cos(e(x),e(x')), where e(·) is the image embedding. If two candidate images come from the same inspection pass and d(x,x')≤0.1, retain only the higher-scoring image unless morphology is distinct. For reproducibility, morphology is treated as distinct when at least one prespecified descriptor differs by a material amount: crack-length-to-box-diagonal ratio by at least 0.2, estimated mean width by at least 0.1 after normalization to image size, branch-count category from unbranched to branched or vice versa, or orientation category among predominantly horizontal, vertical, or oblique. When automatic morphology estimates are unavailable, the lower-scoring image is discarded without exception. This default rule favors consistent filtering over ad hoc retention.
Annotation granularity is chosen by a task-value rule that uses the same Monte Carlo predictions defined above. Send image x to box annotation when localization is uncertain but boundary detail is not: specifically, when the detector dispersion term a_det(x) exceeds the threshold τ_det while the segmentation ambiguity term a_seg(x) is below τ_seg. Here a_det(x) is the mean, over the retained matched detections, of the larger of two quantities: the standard deviation of crack probability across T passes and one minus the mean pairwise intersection-over-union of the corresponding boxes across passes. Let a_seg(x) be the mean pixelwise normalized entropy over the predicted crack region, estimated from the T-pass mean probability map thresholded at 0.5. Send x to segmentation-mask annotation when a_seg(x)>τ_seg and a_det(x)≤τ_det. Send x to dual annotation when both inequalities hold. The default thresholds are the 0.75 quantiles of a_det and a_seg computed on the unlabeled pool at the current round; sensitivity analysis may also examine fixed quantiles from 0.65 to 0.85 in increments of 0.05. If only a detector is available, the routing collapses to box annotation for selected images. If only a segmenter is available, the routing collapses to segmentation-mask annotation. This rule replaces the undefined phrase "detector confidence interval is wide" with an explicit dispersion criterion and avoids assigning both label types by default. [2] [3]
5. Algorithmic workflow and comparator design
The proposed decision sequence is as follows. Step 1: train a seed detector and, when available, a seed segmenter on the initial labeled set. Step 2: run inference on the unlabeled pool and extract uncertainty, embedding, metadata, and morphology descriptors. Step 3: score each image by the acquisition rule and partition the top-ranked images into box, mask, or dual-annotation queues using the same routing rule in every study arm. Step 4: add the newly labeled data to L, retrain or fine-tune the models, and repeat. The stopping rule should be fixed in advance. The default rule is to stop when either the annotation budget is exhausted or the mean held-out improvement over the previous two rounds is less than a prespecified margin ε for two consecutive rounds, with ε defined on the primary metric before data collection. Step 5: compare each active-learning round against a random batch and an uncertainty-only batch of equal annotation cost, not merely equal image count. The primary joint improvement is the equal-weight mean of detector average precision and segmentation intersection-over-union, each on [0,1], measured on a fixed development validation set. If segmentation is unavailable, detector average precision alone is the preregistered primary metric. Stop comparisons use this scalar and epsilon on that scale; the untouched test set never determines stopping.
The comparator set should include at least two baselines: uniform random sampling and uncertainty-only sampling. Random sampling estimates the value of active selection itself; uncertainty-only sampling tests whether the added diversity and morphology terms contribute beyond model confidence. To avoid confounding sample selection with annotation granularity, all comparators should use the same annotation-routing logic and the same unit-cost schedule as the main method after samples are chosen. If computational resources allow, a third comparator can be a fixed-per-bridge quota sampler that selects evenly across inspection sites. That comparator is relevant because bridge data are often clustered by structure and capture session, which may bias ordinary uncertainty sampling toward repeated textures from the same location. [4] [1] [2]
The base models used in comparison should be a bridge-focused detector and a bridge-focused segmenter rather than generic vision models. For detection, improved YOLO-family methods are an appropriate reference class because recent bridge-crack detectors have explicitly targeted slender, low-contrast targets and complex textures. For segmentation, lightweight crack segmenters with fine-detail recovery are the relevant reference class. The purpose is not to duplicate architectural claims, but to test whether annotation policy is still useful once a competent bridge-specific learner is available. [1] [2] [3] [6]
6. Validation design and outcome measures
The prospective evaluation should use temporally separated bridge inspection rounds when possible. A practical design is to train on earlier labeled rounds and then acquire labels from later rounds through the active-learning policy. This reduces leakage from nearly identical frames and better reflects operational inspection practice. If temporal separation is impossible, bridge-wise separation should be used so that images from one bridge are withheld from acquisition training during evaluation of another bridge. [4] [5]
Primary outcomes are annotation efficiency and model quality. The primary estimand for annotation efficiency should be fixed before the study starts: the difference, at a matched annotation cost C, between the held-out performance of the proposed method and that of each comparator. Cost should be measured in annotation units with prespecified weights, for example 1.0 for one box label, 4.0 for one segmentation mask, and 4.5 for one dual label; these defaults may be replaced only by a preregistered time-based schedule obtained from a pilot timing exercise. Model quality should be reported separately for detection and segmentation using task-appropriate metrics such as precision, recall, F1 score, average precision, mean intersection over union, and Dice score. Because bridge inspection often involves class imbalance and small objects, calibration should also be measured using expected calibration error or a comparable reliability statistic. [1] [2] [3] [6]
Secondary outcomes include cross-bridge transfer and stability across random seeds. Cross-bridge transfer is important because the source studies report that bridge condition, imaging setting, and surface texture can alter detection behavior. Stability should be assessed by repeating the same acquisition policy under different training seeds and comparing the variance of the outcome metrics. The result to seek is not a single best score, but a lower labeling burden for the same level of held-out performance relative to the comparators. [1] [3] [5]
7. Failure cases and boundary conditions
The protocol has four main failure cases. First, if the initial seed set is too small or too homogeneous, the uncertainty estimates may be poorly calibrated, and the active learner may repeatedly choose uninformative edge cases. Second, if bridge metadata are unavailable or unreliable, the diversity term may fail to distinguish redundant images from genuinely novel ones. Third, if the image stream contains severe motion blur, overexposure, or heavy occlusion, even carefully selected images may not support either detection or segmentation. Fourth, if the annotation goal shifts toward metrology rather than recognition, box-based sample selection may become insufficient because width measurement requires more exact boundary information. [4] [5] [2]
Generalization is therefore limited to bridge imagery where cracks are visible enough for image-based labeling and where the inspection workflow yields repeatable acquisition conditions. The protocol is not intended for hidden defects, sensor-only monitoring, or non-bridge surfaces. It also does not assume that every bridge type requires a different score. The default operational rule is to use the prespecified weights 0.5, 0.3, and 0.2 for uncertainty, coverage deficit, and redundancy magnitude, respectively. Sensitivity analysis should then vary one weight at a time by δ∈{-0.1,-0.05,0.05,0.1}, reject any perturbation that would move that weight outside [0,1], and rescale the other two weights proportionally so that the three positive weights still sum to 1. Any departure from the default should be chosen by a preregistered tuning rule on a validation split within the seed data rather than adjusted after observing held-out outcomes. [4] [5] [2]
Acquisition and evidential boundaries
The Bayesian deterioration source motivates joining inspections with monitoring; it does not validate active annotation selection. Bridge/session-separated sampling is a proposed defense against dependent images, not an established superiority claim. At each acquisition round, compute score quantiles on the remaining eligible unlabeled pool using that round’s frozen model. Normalize predicted annotation costs to minutes using development-only timing measurements, with a strictly positive minimum cost. Acquisition weights are nonnegative and sum to one after excluding unavailable components. An unavailable novelty representation sends an image to a separately recorded manual eligibility queue rather than assigning zero novelty or silently discarding it. Freeze routing thresholds and budget units before comparison; the same routing rule applies in every comparator arm. Annotation value, stopping margins and weight values are author-proposed settings whose effect requires evaluation, not findings imported from the cited inspection literature.
8. Conceptual contribution and conclusion
This manuscript makes a bounded conceptual contribution. It proposes an active-learning protocol for bridge-crack annotation that combines uncertainty, morphology, and inspection-context coverage in one reproducible sample-selection rule, and it specifies how selected images are routed to box, mask, or dual annotation within detection and segmentation workflows. It also places that proposed workflow within remote bridge inspection and uses hierarchical Bayesian deterioration modelling as an interpretive framework for thinking about underrepresented crack regimes and inspection uncertainty. [4] [5] [1] [2]
What remains untested is whether the proposed score and routing rule outperform uncertainty-only or random selection across bridge types, capture conditions, and labeling budgets. The appropriate next step is a controlled prospective study with temporally or spatially separated bridge data, prespecified annotation costs, a fixed stopping rule, and repeated-seed evaluation. Until such evidence is collected, the protocol should be regarded as a precise and falsifiable method specification rather than as an empirically validated claim of superiority. [1] [2] [3] [5]
Contribution and availability
A conceptual research protocol that specifies a reproducible active-learning framework for bridge crack annotation, including normalized uncertainty, redundancy, and coverage terms; a default weighted acquisition rule with sensitivity analysis; explicit annotation-routing logic shared across comparators; prespecified efficiency estimands and stopping rules; and a prospective evaluation design grounded in bridge inspection practice and probabilistic deterioration modelling.
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
- 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.
- 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.
- 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.
- 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.
- Wang, Benyu; Chen, Ke; Wang, Bingjian (2026). Hierarchical bayesian fusion of inspection and monitoring data for probabilistic bridge deterioration assessment. Scientific Reports. 16(1), 5965. doi:10.1038/s41598-026-36808-4.
- 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.