Multiple Feature Fusion Based on Hierarchical Constraint for Crack Detection

Because of their security and convenience, crack detection methods based on image processing technique (IPT) have gradually become a mainstream trend in this field. However, the performances of existing methods can not satisfy the practical applications for serious category imbalance and uncontrolled noisy interferences. In this paper, a multiple feature fusion framework based on hierarchical constraint for crack detection is proposed, which mainly contains a weight-shared encoder component and three independent decoder ones, and the outputs of these three decoder components are fused together to finalize the detection. Specifically, each decoder component is constrained by pseudo label generated from the true crack map with a specific expansion stride, which can alleviate the serious category imbalance problem and avoid the risk of model collapse. Furthermore, the fused feature can represent the road image well because it can ensure the existence and meanwhile avert the false-alarm of crack samples. In addition, the backbone of our framework is based on a lightweight UNet equipped with enhancement modules, which can increase the structural representation capability, decrease the computational load, and speed up the reasoning time simultaneously. Finally qualitative and quantitative experimental results on three challenging datasets have verified the superiority of the proposed framework.