DetCrack: concrete crack detection algorithm based on single-stage object detection model

Surface cracks in concrete structures can severely impact the safety and stability of buildings. To achieve efficient concrete crack detection, this paper proposes a single-stage concrete crack detection model, DetCrack, composed of a fundamental feature extraction network, a multi-level feature fusion network, and three detection heads. The usage of dynamic snake convolution (DSConv) allows for adaptive adjustment of the convolution kernel’s size and shape, precisely capturing crack features and enabling the model to better adapt to the diversity and complexity of cracks. The DetCrack model is evaluated, and the average precision AP_0.5/%, AP_0.5-0.95/%, number of parameters and computational cost are 92.8%, 76.7%, 1.71M, and 5.0GFLOPs, respectively, surpassing comparative methods. The experimental results indicate that the DetCrack model possesses robustness and is a highly accurate, lightweight model for concrete crack detection.