| Title | Authors | Year | Why it’s relevant | |-------|----------|------|-------------------| | DeepCrack: Learning Hierarchical Features for Crack Detection | Z. Zhang, et al. | 2019 | Demonstrates how CNNs can outperform classic edge detectors when trained on the Carlson dataset. | | UAV‑Based Photogrammetric Crack Mapping of Pavements | S. M. Lee, J. Kim | 2021 | Extends Carlson’s ground‑based method to aerial platforms; includes flight‑planning guidelines. | | Robust Crack Width Measurement Using Structured Light | M. B. G. Alvarez, et al. | 2022 | Compares active‑projection techniques to passive photography; useful for very narrow cracks (< 0.1 mm). | | Uncertainty Propagation in Image‑Based Structural Health Monitoring | L. P. Huang, et al. | 2023 | Provides a modern Bayesian framework that can be applied on top of Carlson’s deterministic error budget. |
| # | Requirement | Acceptance Criteria | |---|-------------|---------------------| | | Capture‑time preview – When a user takes a photo, the UI overlays a quick crack‑heatmap (low‑resolution) within 500 ms. | Users see a translucent red overlay that disappears once the full analysis finishes. | | FR‑2 | Full‑resolution analysis – Run a high‑accuracy model on the saved image and produce a detailed mask. | Mask aligns pixel‑perfectly with the original; processing time ≤ 2 s for 12 MP JPEG on a GPU‑enabled server. | | FR‑3 | Crack metrics – For each detected crack, compute: • Length (mm) • Maximum width (mm) • Average width (mm) • Orientation (°) • Bounding box & polygon. | Metrics appear in a scrollable “Crack List” UI and are exportable as JSON/CSV. | | FR‑4 | Severity scoring – Map metric ranges to a 1‑5 severity level (or custom thresholds). | Example: Level 1 = width < 0.2 mm, length < 20 mm Level 5 = width > 2 mm or length > 200 mm. | | FR‑5 | Export / API – Provide: • JSON payload per image • Annotated image (original + mask overlay) • CSV batch export. | External systems can pull /api/v1/crack‑detect/imageId and receive the payload. | | FR‑6 | User feedback loop – Users can “Accept”, “Reject”, or “Edit” a detected crack. Rejected masks are stored for future model fine‑tuning. | A “thumbs‑up/down” UI element next to each crack; rejected items are flagged in the data lake. | | FR‑7 | Offline fallback – On devices without connectivity, run a lightweight TensorFlow‑Lite model locally and sync results later. | The same UI works; sync status is shown in a “Pending Upload” queue. | | FR‑8 | Access control – Only users with the role Inspector or higher can view raw masks; other roles see only scores. | Role‑based UI component hiding verified in unit tests. | | FR‑9 | Audit trail – Every analysis run logs: user‑id, timestamp, model version, hardware (GPU/CPU), and processing duration. | Logs are searchable via /admin/audit . | | FR‑10 | Performance monitoring – Emit Prometheus metrics: ccd_processing_seconds , cdd_detected_cracks_total , cdd_false_positives_total . | Grafana dashboard alerts if latency > 3 s for > 5 % of requests. | carlson photo capture crack
In essence, Carlson PhotoCapture does not just document damage; it provides a framework for . By digitizing the physical flaws of the world, it allows us to analyze the "crack" in safety, far away from the risk of the structure itself. standalone processing? | Title | Authors | Year | Why