Deep learning-based multi-category disease semantic image segmentation detection for concrete structures using the Res-Unet model

This paper presents an enhanced Unet network (Res-Unet) for the identification of prevalent concrete diseases, namely cracks, spalling, holes, alkaline flooding, and exposed reinforcement. The proposed approach involves constructing a Res-Unet network model by integrating the Unet model with the ResNet50 network. Additionally, the training dataset is augmented using geometric deformation techniques, such as cropping, rotating, and mirroring, applied to the original images depicting common concrete diseases. In this study, a total of 13,200 concrete illness picture datasets were acquired. These datasets were used to train, validate, and test several models, including the Res-Unet model, the original Unet model, the VGG16 + Unet model, the FCN + VGG16 model, and the FCN + ResNet50 model. The findings indicate that the Res-Unet model achieves a mean Intersection over Union of 86.3% and an average pixel accuracy of 98.5%. The improved Res-Unet model ensures the accurate extraction of the whole fracture skeleton and the identification of minor cracks. This study's findings can be utilized to identify specific damages in concrete in real-world engineering scenarios precisely.