Dual-Branch Feature Fusion Network for Concrete Crack Depth Prediction

Concrete structures are prone to cracking, particularly with crack depth significantly affecting their structural integrity and safety. Traditional methods for estimating crack depth are labor-intensive and highly subjective, highlighting the need for automated detection methods. Existing deep learning models for automatic detection focus primarily on estimating crack width, length, and area. Therefore, this study proposes a multibranch deep learning model, the Dual-Branch Feature Fusion Network (DBFFNet), for crack depth estimation. The model utilizes a ResNet-based approach to extract high-level features from red-green-blue images. Due to the pronounced edge features in binary crack images, convolutional blocks are used to extract shape and boundary features. The features from both branches are enhanced using a channel attention module and further processed through convolutional block group layers. The enhanced features are fused through various operations and processed by a Multi-Input Global-Local Attention Fusion module to improve feature expression accuracy and completeness. The final depth prediction is performed by a multilayer perceptron. Experimental results demonstrate that DBFFNet outperforms the latest image-based crack depth prediction methods and traditional methods in terms of mean squared error and correlation coefficient (R2) metrics, providing a reliable tool for evaluating crack depth and ensuring structural integrity.