Vision based crack detection and growth measurement in mechanical structures
Azeem, Muhammad Hamza (2026)
Kandidaatintyö
2026
School of Energy Systems, Konetekniikka
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2026052049752
https://urn.fi/URN:NBN:fi-fe2026052049752
Tiivistelmä
This bachelor’s thesis examines vision-based crack detection and crack growth measurement in mechanical structures. Crack estimation is important in fatigue-loaded structures because visible cracks indicate damage development that may reduce structural safety and service life. Conventional visual inspection is often slow, subjective, and difficult to repeat consistently, especially when cracks are thin and appear in the presence of welds, shadows, reflections, and other crack-like surface features. For this reason, a more reliable and automated vision-based method is needed for crack identification and measurement.
To address this problem, the study proposes a two-stage framework for automated crack assessment. First, specimen images are collected, converted into a trainable format, and labelled using a semi-automatic annotation tool. A deep-learning-based object detector is then used to localise the crack region in the full image. In the second stage, classical image processing methods are applied inside the detected region of interest in order to segment the crack more precisely and estimate its growth relative to a predefined reference length.
The developed workflow is evaluated using numerical detection metrics. The trained model achieved a best F1-score of 72.1% at a confidence threshold of 50.1%, and an mAP@0.5 value of 71.7%. In addition to these quantitative results, qualitative examples showed that the method could produce interpretable outputs in the form of crack localisation, and percentage-based crack growth estimation.
The results indicate that the combination of deep learning and classical image processing provides a practical approach for visible fatigue crack monitoring. The study concludes that vision-based crack inspection has clear potential to support structural health monitoring and maintenance-related decision-making, particularly when the inspection task requires not only crack recognition but also a measurable description of crack development.
To address this problem, the study proposes a two-stage framework for automated crack assessment. First, specimen images are collected, converted into a trainable format, and labelled using a semi-automatic annotation tool. A deep-learning-based object detector is then used to localise the crack region in the full image. In the second stage, classical image processing methods are applied inside the detected region of interest in order to segment the crack more precisely and estimate its growth relative to a predefined reference length.
The developed workflow is evaluated using numerical detection metrics. The trained model achieved a best F1-score of 72.1% at a confidence threshold of 50.1%, and an mAP@0.5 value of 71.7%. In addition to these quantitative results, qualitative examples showed that the method could produce interpretable outputs in the form of crack localisation, and percentage-based crack growth estimation.
The results indicate that the combination of deep learning and classical image processing provides a practical approach for visible fatigue crack monitoring. The study concludes that vision-based crack inspection has clear potential to support structural health monitoring and maintenance-related decision-making, particularly when the inspection task requires not only crack recognition but also a measurable description of crack development.