Arman NeyestaniComputer Vision · MeasurementResume
← All work

Structural Monitoring

UAV-assisted concrete crack segmentation

Pixel-level crack masks for civil-infrastructure inspection, developed around a YOLOv8 segmentation workflow and UAV-acquired imagery.

Code repository ↗IEEE paper ↗
My role
I led the study as first author, preparing imagery and annotations, setting up data splits and augmentation, training YOLOv8 segmentation variants, and evaluating failure cases.
What I built
A documented training entry point and portable inference script accept images, folders, video, or streams. The repository separates code from large research artifacts.
Key results
  • 0.62–0.64 F1 score across five model sizes (522 held-out test samples)
  • 0.85–0.87 average precision across five model sizes (published comparison)
Stack
Python · Ultralytics YOLOv8 · PyTorch · OpenCV · Polygon annotation · UAV imagery
Examples of predicted and reference crack masks from the UAV inspection study.
Prediction and ground-truth mask examples from the published study.

Problem

What needed to be understood or measured

A detection box says a crack is present; an inspection workflow often needs its pixel-level extent. This study evaluated instance segmentation as a path from survey imagery to reviewable crack masks.

My contribution

My contribution

I led the study as first author, preparing imagery and annotations, setting up data splits and augmentation, training YOLOv8 segmentation variants, and evaluating failure cases.

Experimental setup

Data and setup

Civil-infrastructure images with annotated crack regions; YOLOv8 segmentation variants trained for 20 epochs at 448 px and evaluated on a held-out split.

Six concrete crack images comparing predicted red masks with green ground-truth boundaries.Open figure —
Publication figure

Six crack-segmentation examples. Red is the model prediction; green is the ground-truth boundary.

Concrete Crack Detection and Segmentation in Civil Infrastructures Using UAVs and Deep Learning, Fig. 1, p. 3.

Method

From input to interpretable output

A documented training entry point and portable inference script accept images, folders, video, or streams. The repository separates code from large research artifacts.

Results

Reported results

0.62–0.64

F1 score across five model sizes

522 held-out test samples

0.85–0.87

average precision across five model sizes

published comparison

YOLOv8 medium model training, validation, precision, recall and mean average precision plots.Open figure —
Publication figure

Training and validation traces for the medium YOLOv8 segmentation model.

Concrete Crack Detection and Segmentation in Civil Infrastructures Using UAVs and Deep Learning, Fig. 2, p. 4.

Validation & uncertainty

Reference and uncertainty checks

I compared five YOLOv8 sizes on a held-out set of 522 samples. Across the models, reported F1 scores range from 0.62 to 0.64 and average precision from 0.85 to 0.87. The qualitative masks show that predictions can extend beyond the annotated crack boundaries.

Limitations & failure modes

Operating limits and failure modes

Fine cracks, surface texture, shadows, low contrast, motion blur, viewpoint changes, and annotation ambiguity can produce missed or fragmented masks.

The public repository does not include source data or trained weights. It is reproducible code, not a one-click hosted model or evidence of field deployment.

Technical stack

Tools selected for the measurement chain

  • Python
  • Ultralytics YOLOv8
  • PyTorch
  • OpenCV
  • Polygon annotation
  • UAV imagery

Sources & project links

Papers and code

IEEE IoT 2023 paper and public repository. Repository figures retain their publication attribution; no dataset or weights are redistributed here.

Code repository ↗IEEE paper ↗