A High-accuracy Crack Defect Detection Based on Fully Convolutional Network Applied to Building Quality Inspection Robot

Crack defect detection plays a significant role in the industry, especially in construction. This paper proposes a high accuracy detection method based on the improved fully convolutional network (FCN), which can be used in the building quality inspection robot. After building the map, an algorithm based on topology is applied to recognize the contour of the house, plan the path and take photos. The up-sampling method is improved based on the existing network structure, and a crack detection model is built. A Deepcrack dataset is used to train the model and tested with 100 images in various scenes. Experimental results with mean-IoU of 48.27% and accuracy of 92.81% indicate that the proposed detection model can accurately detect the cracks at the pixel-wise level.