MRSA-U-NET: A Novel Deep Learning Approach for Crack Segmentation in Structural Health Monitoring

Crack segmentation is a very important part of the Structural Health Monitoring System.The SNI, ACI 224R-01, and Eurocode 2 standards set different maximum crack width, but no more than 4 mm for various types of buildings and environmental conditions.Crack segmentation usually focuses on optimizing objective functions and ignores the presence of small cracks.To address this issue, we are introducing a new architecture that called Multi-Resolution Spatial Attention U-Net (MRSA-U-NET).MRSA-U-NET has three main components, namely Multi-Scale, Residual Connections, and Convolutional Block Attention Module (CBAM).Multi-Scale in this architecture uses Atrous Spatial Pyramid Pooling (ASPP) to capture features at different scales without reducing spatial resolution.Residual Connections are used to improve network learning by using skip connections to avoid degradation in deep networks.Finally, Spatial Attention is one part of CBAM to focus on important spatial locations.The results show that the effectiveness of the designed new architecture Multi-Resolution Spatial Attention U-Net (MRSA-U-NET) can overcome the challenges in detecting and segmenting small cracks The superior F1 Score and IOU values in all datasets indicate that this method can be applied to various types of cracks.These results represent a significant step toward crack detection systems that meet the maximum crack width standards.