Thesis
Adaptive lighting based detection of cracks and spalling in concrete structures
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17354
- Person Identifier (Local)
- 202054421
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Defects such as cracks and spalling in reinforced concrete can have a detrimental effect on structural integrity if unnoticed and left untreated. Inspection methods have largely evolved over the years from traditional visual inspections to partially and fully automated inspections. One of the most common forms of automated inspections is the use of cameras and application of deep learning algorithms to identify cracks and spalling on the captured images. Automated inspections overcome limitations of visual inspections, mainly as it concerns inconsistency and subjectivity in the identification and interpretation of cracks. Still, they do not come without their own caveats: autonomous vehicles (e.g. drones) require an expert, trained and licenced user, the captured images use diffused lighting that does not necessarily reveal the true geometric characteristics of the crack or even the crack itself, and most deep learning algorithms developed for crack detection and identification are not performing well with low quality images and low light conditions and are mainly trained with diffused light images captured in the lab. This thesis is focusing on addressing all of these challenges in the technology of automated concrete structure inspections in Civil Engineering through the development of both hardware and software. Utilising the benefits of multi-directional, multi-angle lighting, which has been successfully used in other disciplines (medicine) in highlighting features clearly, an automated platform called ALICS (Adaptive Lighting for Inspection of Concrete Structures) has been developed. It captures images of concrete structures by creating and casting shadows through illuminating light from multiple angles and multiple directions, mimicking what an experienced inspector would do during a visual inspection. The original design of ALICS has subsequently been modified to field-deployable hardware, suitable for carrying out inspections outside the lab. Following a number of experimental setups where the angle of light was varied and analysis of the captured images using a pretrained VGG-16 neural network model, the optimum angle was identified. To overcome declined performance of the VGG-16 model, one of the most commonly used models in concrete crack identification in Civil Engineering, in the case of low-quality images, an automated image quality assessment workflow was developed based on BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) score and incorporated into the VGG-16 model. The traditional VGG-16 and VGG-19 neural network models are developed to satisfy three-channel diffused RGB images. To allow for the analysis of directional lighting, i.e. for five channels, one for the gray scale image in Right, Down, Left, Up and diffused directions each, novel five-channel VGG-16 and VGG-19 were developed, capable to detect and classify cracks. The VGG-16 five-channel neural network was further improved in terms of evaluation time by utilising the light neural network model, MobileNetV2. The multi-channel MobileNetV2 model, MCNet has been implemented similar to five-channel VGG models. The maximum-intensity fusion technique, which has been successfully used in medical field to combine multiple images has been utilised to implement a fused neural network model, FusedNN. The FusedNN, and MCNet models are developed to satisfy three-channel RGB images, and five-channel images, respectively. The performance ofthe traditional MobileNetV2 model, FusedNN, the five-channel VGG-16 model, and MCNet were compared for crack detection and classification tasks, and MCNet demonstrated the best performance. For detecting cracks and spalling, a comparison of the four models: traditional, Zoubir, FusedNN, and MCNet, revealed that MCNet outperformed the others. This highlights the advantage of additional information provided by directional lighting to enhance automated concrete crack inspection in Civil Engineering. The performance of traditional model and the FusedNN model was compared under increased exposure values to evaluate the advantages offered by the FusedNN model over traditional model. Overall, this research offers a comprehensive approach to automated concrete inspection, leveraging, for the first time in Civil Engineering, advanced illumination technologies used successfully in other scientific fields. The results in this thesis show that multi-directional lighting, combined with specifically adapted neural networks can successfully identify cracks of widths 0.1mm.
- Advisor / supervisor
- Pytharouli, Stella
- Dobie, Gordon
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of Thesis T17354 | 2025-07-16 | Public | Download |