Deep Learning Networks for Concrete Crack Image Segmentation

Cracks in concrete pose a significant threat to building safety, making the detection and identification of these cracks essential to ensuring occupant safety. Currently, concrete crack detection is mostly conducted manually, which can introduce subjectivity, inaccuracies, and inefficiency. To address these issues, this paper proposes five deep learning models to replace manual detection of concrete cracks: the Pyramid Scene Parsing Network(PSPNet), which combines global and local information to improve semantic segmentation accuracy; the Dilated Dense Residual Network(DDRNet),which utilizes a bilateral fusion strategy to integrate information from different feature maps for fast and accurate segmentation; the Fully Convolutional Network(FCN), which simplifies the network structure to ensure that input and output dimensions are the same; the Artificial Neural Network(ANN), which enhances the model’s ability to recognize objects at different scales through multi-scale fusion; and the Criss-Cross Network(CCNet), which employs a recursive cross-attention module to reduce computational load and improve segmentation accuracy. Experimental results demonstrate that the IoU of these five networks is 43.63% or higher, and the F1 score is 59.36% or higher, offering a novel approach for concrete crack detection.