Bi-DAUnet: Leveraging BiFormer in a Unet-like Architecture for Building Damage Assessment

Abstract

In recent years, Convolutional Neural Networks (CNNs) have become an important research direction in the field of building damage assessment. Particularly, deep neural networks based on the U-shaped architecture and skip connections have achieved significant breakthroughs in the task of architectural damage assessment. Despite the impressive performance of CNNs, effectively capturing global and long-range semantic information remains a challenge due to the local nature of their convolutional operations. To address this issue, we propose a novel architectural damage assessment model called Bi-DAUnet, which adopts a BiFormer structure similar to U-Net. In this model, we employ a U-shaped encoder-decoder architecture based on BiFormer and combine it with skip connections to achieve global semantic feature learning. Specifically, we utilize a hierarchical BiFormer with a dual-layer routing attention mechanism as the encoder to extract contextual features of architectural images. In the symmetric decoder, a BiFormer Block is introduced to fuse shallow and deep features of the feature maps and learn the correlation between pixels at distant locations. Experimental results indicate that the U-shaped encoder-decoder network based on BiFormer achieves superior performance in the task of architectural damage assessment compared to fully convolutional methods.