Published Paper Details:

In civil engineering, the structural integrities of bridges are very important. This is because problems with the bridge surface can make bridges unsafe, more costly to maintain, and even lead to catastrophic failure. Most traditional methods for inspecting bridges are manual, where trained personnel look for cracks, spalling, corrosion and other issues on the surfaces of. These manual inspections work to some extent, but they are time-consuming and labour-intensive and are prone to error. Moreover, manual assessments are subjective and often do not agree with each other, especially when there are multiple defects or when surface conditions are made more difficult by lighting, noise, or other environmental factors. In recent years, machine learning and deep learning have emerged as powerful tools to automate the detection of defects in civil infrastructure. Among them, convolutional neural network (CNN) based object detection models like the YOLO (You Only Look Once) family have shown a lot of promise for real-time detection tasks due to their speed and efficiency. But the models that are already out there have some issues. Many of them are designed to find only one kind of defect, making them less useful in the real world where multiple defects often happen at the same time. Standard models also lack mechanisms to focus on the most important parts of an image, making it more difficult for them to detect small, subtle or overlapping defects. Bounding box regression uses standard loss functions such as IoU or GIoU but these may not be effective with defects that are not in the shape of a box which can lead to errors in localisation . These limitations highlight the necessity of a more powerful and precise system capable of detecting various kinds of defects on bridge surfaces under different challenging and variable circumstances. We propose a YOLOv8 model with an improvement using the Convolutional Block Attention Module (CBAM) and the Wise-IoU loss function in this paper. This new model is named YOLOv8-CBAM-Wise-IoU. The CBAM module enhances the feature representation by employing channel and spatial attention. This guides the model to focus on relevant portions of the image and ignore irrelevant background noise. This is even more important for detection of small cracks, subtle corrosion or overlapping defects, that could be missed by standard detection models. The Wise-IoU loss function improves bounding box regression by dynamically adjusting the loss weight according to the object characteristics. This leads to more accurate localisation and reduces false positives. The proposed model can detect seven different types of defects on bridge surfaces simultaneously. They consist of cracks, spalling, corrosion, delamination, surface scaling, efflorescence and potholes. We collected a large set of labelled bridge pictures for training the model, and augmented it with methods such as rotation, scaling and brightness changes to make it more generalisable. We put a lot of effort into improving the training, tuning hyperparameters, performing ablation studies to evaluate the contribution of each component (YOLOv8 backbone, CBAM, and Wise-IoU) to the overall performance. Tests show that YOLOv8-CBAM-Wise-IoU performs much better than the standard YOLO models and other common methods. The model could detect 97.9% of the defects, 76% of the times it was supposed to, 58% of the times it was supposed to and 55.4% of the times it was supposed to. It was also able to detect defects in real world scenarios such as variations in lighting, complex backgrounds and non-uniform surface textures. The system also had a faster detection speed which was good for real time use and this meant it could be used for inspection systems that use drones or cameras. In summary, the YOLOv8-CBAM-Wise-IoU model is a reliable, scalable and efficient method for multi-defect detection on bridge surfaces. To solve the problems of traditional detection methods, the system adopts the attention mechanism and more advanced loss function. Enables the complete, accurate and automated monitoring of the health of structures. It could enhance preventative maintenance schedules, lower the cost of inspections and make bridges safer overall. This study adds to the development of intelligent infrastructure inspection systems, and also shows the effectiveness of combining deep learning, attention mechanisms, and optimised loss functions to address practical civil engineering applications.