Deep Learning Image-Based Fusion Approach for Identifying Multiple Apparent Diseases in Concrete Structure

Addressing the key pain points in detecting typical apparent diseases of concrete structures, where standalone object detection fails to achieve pixel-level quantification and standalone semantic segmentation, is inefficient. Therefore, a deep learning image-based fusion approach is proposed to identify the typical visible diseases in concrete structures, namely crack, spalling, water leakage, and seam deformation. To implement the approach, a deep learning fusion network is developed with the YOLO and UNet models to identify multiple apparent diseases rapidly. In the fusion network, the YOLO model is used to filter the images containing the visible diseases from all the images in the first stage. Then, the UNet model is used to extract the pixels containing diseases from the selected images. Lastly, analysis methods are proposed to quantify the diseases based on the segmented pixels, such as length, width, and area. In this paper, a dataset of 1488 images with the above diseases from a field inspection was used to train the deep learning fusion network. The training results demonstrated the robustness of the fusion network in identifying and segmenting diseases with a mean average precision of 0.72 and a Dice score of 0.82. Experiments were finally conducted on concrete slabs with simulated diseases for additional validation. The results indicated that the proposed fusion network could identify the diseases approximately 50% faster than the UNet model only. The quantification precision was found to be satisfactory, with relative errors below 11.07% for the area of water leakage, below 5% for the length and area of cracks, and below 6% for the width of seams.