A review and analysis of neuromorphic computing technology to detect concrete structure defects using object detection

Bourke, Allan (2023) A review and analysis of neuromorphic computing technology to detect concrete structure defects using object detection. [USQ Project]

[img] Text (Project – redacted)
Bourke_A_Nguyen_Redacted.pdf

Download (6MB)

Abstract

Deep learning utilises neural network layers to systematically process similarities and differences between images to establish a reliable set of features used to assist in localising and classifying objects within an image. This is done through an object detection model and the output is an application of a class label to the object within the image. Object detection models could automate, and therefore greatly expedite, the inspection of concrete structures which are currently periodically visually inspected by trained inspectors to determine asset condition and maintenance and requirements. This manual process is considered ‘Gold Standard’; however it is time-consuming, expensive, and triggers field hazards. To date, many research studies have demonstrated successful use of second-generation artificial intelligence (AI) technology to identify defects in concrete structures, attempting to replicate the inspector. More recently, advances in AI systems have led to the development of third-generation neuromorphic computing technology and its use for object detection has the potential to increase detection speed and efficiency, use less power, enhance security, and allow for data to be analysed locally without requiring large cloud-based data stores. The overarching aim of this project was to demonstrate that neuromorphic computing technology is a suitable novel technology to detect common defects on concrete bridge and culvert structures using object detection. Specifically, the project aimed to develop, train and implement a neuromorphic computer vision model to identify common bridge defects and determine the system accuracy, effectiveness and usability. The model was also directly compared with a traditional object detection model.

Photographic images (n = 844) of concrete structures were manually collected through field inspections and dissected into 17395 512 x 512 pixel images. These were manually classified to obtain 1326 images of various structure defects. Of these, 200 high quality crack and spall images were utilised to develop YOLOv5 and AKIDA object detection models using Edge Impulse studio with three hundred (300) training cycles. The model configuration was set using a bounding box labelling method. For all models, 64% of images were allocated to training, 16% for validation, and 20% for model testing. Crack only models were developed using 100, 150, 170, 200, 350, 300 and 350 images and a spall and crack model using 100 spall and 100 crack images was also developed. The learning rate varied between 0.01 and 0.001 for AKIDA to optimise testing results while the learning rate of 0.01 for the YOLO5 model could not be modified through the online portal. Confidence thresholds of 30%, 50% and 70%t were set to analyse the accuracy of these models using the test dataset. Finally, the performance of both models was visually assessed in real-time environment replicating a field setting.

Both models produced poor results overall for the combined crack and spall image dataset, with less than 30% accuracy using the lowest confidence threshold. Further, they produced 0% accuracy for spalls using all three (3) confidence thresholds specified. However, both models were able to successfully detect concrete cracks in the images. The AKIDA model produced the highest precision (94.1%) of all models on the validation data, using 150 images, while YOLO only achieved 69%. The highest precision able to be obtained with the YOLOv5 model was 75% using 350 crack images. For accuracy, the AKIDA model achieved 81.7% accuracy with 300 images using the 30% confidence threshold and 68% accuracy under the 70% threshold. The highest accuracy achieved by YOLOv5 was 25% using 200 images. Conversely, the YOLO model performed better in the visual inspection test, with more accurate crack detection in 10 out of 14 instances through this subject assessment process.

The study highlighted that image quality can negatively affect results and model development. In particular, neither model was able to successfully detect spalls, as these are a more complex defect, making them more difficult for the models to detect. However, both models were able to detect cracks, with the AKIDA model demonstrating that third-generation models were comparable to second-generation technologies for the identification of cracks in this setting. Further investigation could incorporate additional defects into the object detection model, to simulate a comprehensive structure inspection. The study suggests neuromorphic computing is a new promising technology to identify common defects in concrete structures.


Statistics for USQ ePrint 52933
Statistics for this ePrint Item
Item Type: USQ Project
Item Status: Live Archive
Faculty/School / Institute/Centre: Current – Faculty of Health, Engineering and Sciences - School of Engineering (1 Jan 2022 -)
Supervisors: Nguyen, Andy; Brown, Jason
Date Deposited: 23 Sep 2025 03:29
Last Modified: 23 Sep 2025 03:29
Uncontrolled Keywords: computing technology; AKIDA
URI: https://sear.unisq.edu.au/id/eprint/52933

Actions (login required)

View Item Archive Repository Staff Only