Efficient attention-based deep encoder and decoder for automatic crack segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Structural Health Monitoring

Background:

• Deep convolutional neural networks have been used for crack segmentation, but face challenges with ground truth data, complex scenes, and object-specific networks.
• Existing methods often struggle with real-time processing and robust evaluation in diverse environments.

Purpose of the Study:

• To develop a novel Semantic Transformer Representation Network (STRNet) for efficient and accurate pixel-level crack segmentation in complex, real-world scenes.
• To address limitations in current crack segmentation techniques, including data preparation and network design for complex scenarios.

Main Methods:

• Developed STRNet, featuring a squeeze and excitation attention encoder, multi-head attention decoder, coarse upsampling, focal-Tversky loss, and a learnable swish activation function.
• Proposed a method for evaluating image scene complexity.
• Trained the network on 1203 images with extensive data augmentation and tested on 545 images.

Main Results:

• STRNet achieved high performance metrics: 91.7% precision, 92.7% recall, 92.2% F1 score, and 92.6% mean intersection over union (mIoU).
• The network demonstrated the fastest processing speed at 49.2 frames per second compared to advanced networks like Attention U-net, CrackSegNet, Deeplab V3+, FPHBN, and Unet++.
• STRNet exhibited superior performance across all evaluation metrics against recently developed advanced networks.

Conclusions:

• STRNet offers a significant advancement in real-time crack segmentation for complex scenes, outperforming existing state-of-the-art methods.
• The proposed network provides a concise and efficient solution for structural health monitoring and infrastructure inspection applications.