Concrete Crack Detection Method Based on Deepcrack Network
Concrete structural cracks pose a great potential threat to building safety,and crack detection is of great significance to the ma-intenance of building structures.The current deep learning-based crack detection for extracting crack details still needs to be improved.Therefore,we optimize the Deepcrack network and propose a concrete crack detection algorithm PG-Deepcrack based on pyramid split at-tention and global context.Firstly,a dual-convolution-attention parallel block is proposed in the encoder to add a pyramid-split attention branch to provide richer multi-scale crack information for the convolutional layer.Secondly,in order to capture long-distance dependencies,a global context block is introduced after the operation of the parallel block,which further improves the ability of network to express the crack details.Finally,the omni-dimensional dynamic convolution and the GELU activation function are utilized in the feature fusion stage to cascade-level fusion of codec features,so that the network retains the information of different sizes of cracks in a more comprehensive way and improves the generalization performance of the model.To validate the effectiveness of the network model,a comparative test is conducted with seven network models on the DeepCrack dataset,and the proposed network exhibits the best performance with an IoU of 72.78％.
万方数据
维普
