AI Bridges Gap Between Poor-Quality Images and Crack Detection Accuracy
Researchers have demonstrated that combining multiple datasets can train AI models to detect concrete cracks in bridge images even when individual datasets are small or low-quality. The finding addresses a real operational problem: inspectors often work with imperfect photos, yet current systems struggle with limited training data—a barrier that has slowed adoption of automated inspection in infrastructure maintenance.
Originaltitel: Concrete Crack Detection Using Multi-Source Data Augmentation in Deep Learning Models
Image processing tasks have benefited from deep learning models based on convolutional neural networks. However, the success of image classification models is dependent on several factors such image quality, dataset size and class distribution. Achieving acceptable accuracies with datasets not meeting these requirements is challenging. Domain specific dataset augmentation techniques have been proposed to mitigate the problem. This paper investigates adaptation of multi-source datasets as an augmentation approach to improve accuracy of crack detection in bridge concrete structures from low quality images in limited and imbalanced datasets. While experimental results show that data augmentation can improve accuracy of detection, we anticipate achieving even better results by combining this approach with generative machine learning models in future research.