Dmg2Former-AR: Vision Transformers with Adaptive Rescaling for High-Resolution Structural Visual Inspection

Area of Science:

• Structural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Drones and advanced imaging enhance structural assessments but image processing is time-consuming.
• Current methods often lose critical details from high-resolution images during processing.
• Efficient and accurate damage detection is crucial for timely structural condition assessments.

Purpose of the Study:

• To develop a semantic segmentation architecture for rapid and accurate pixel-level damage detection in structural inspections.
• To address the challenge of processing high-resolution inspection images efficiently.
• To preserve critical details like microcracks and edges during image analysis.

Main Methods:

• Integration of vision transformers with Laplacian pyramid scaling networks.
• Development of two non-uniform image rescaling networks to reduce computational demands.
• Proposal of Dmg2Former, a low-resolution segmentation network with a Swin Transformer backbone.

Main Results:

• The proposed architecture enables detailed damage identification on high-resolution images.
• Significant reduction in computational demands for image processing.
• Successful validation on public datasets for tasks like crack detection and material identification.

Conclusions:

• The developed method offers a computationally efficient solution for detailed visual inspection analysis.
• Preservation of fine details like microcracks is achieved through adaptive rescaling.
• The approach significantly speeds up the structural condition assessment process using drone imagery.