Research on Non-Destructive Testing of Curtain Wall Hangings Based on Efficient Channel Attention Vision Transformer Network

Abstract

Traditional inspection of curtain wall metal hangings usually relies on manual visual inspection, which is costly, slow, and limited in coverage. To reduce cost and improve efficiency and accuracy, a non-destructive automatic inspection system for architectural curtain walls based on millimeter wave imaging was designed. The system is designed as a single-side reflective point-frequency imaging, which can effectively solve the problem of reflective waves generated on the upper and lower surfaces of the curtain wall during millimeter-wave penetration affecting the echo signal. To address the fine-grained classification problem of metal hangings, we propose an efficient channel attention Vision Transformer (ECA-ViT) lightweight classification network based on a hybrid architecture of convolutional neural network and Transformer. Among them, the inverted residual attention module (IRAM) improves the network’s attention weight on the image foreground, and the low-rank MobileViT module (LR-ViT) can provide global modeling for the network and making the whole model more lightweight by reducing the computational complexity of the self-attention mechanism. The experimental results demonstrate that the proposed method achieves an accuracy of 95.66% with fewer model parameters and computational complexity, demonstrating good performance advantages.

Keywords

• Curtain wall metal pendant
• deep learning
• ECA-ViT
• IRAM