Crack Segmentation of Underwater Structures of Bridges Based on Hierarchical Feature Residual Neural Network
A crack detection method based on hierarchical residual neural network is proposed to improve the automation of the crack detection task for underwater structures of bridges.The method utilizes a multi-level feature residual linkage mechanism,which suppresses the interference of noise features on the building surface,extracts and fuses feature images at different levels,and enhances the model's capacity to accurately delineate cracked and non-cracked regions.Meanwhile,with the help of transfer learning method,the model is initialized with the parameters of the pre-trained model and the weights are adjusted with the underwater crack dataset,so that the model has the capacity to analyze the bridge underwater structure crack dataset with very small amount of data.The model was experimentally validated on a self-collected bridge underwater structural crack dataset.The results showed that the hierarchical residual neural network could accurately classify cracked pixels from non-cracked pixels,and the predicted pixel accuracy reached 87.2%,which proved the feasibility of the method.The model provides an effective solution for automating the bridge underwater structural crack detection task,and also provides a reference idea for other similar image detection tasks.
NETL
NSTL
万方数据
维普
