A Residual-Self-Attention Fusion Network for Identification and Classification of Apparent Defects in Fair-faced Concrete

Abstract

This paper proposes a fair-faced concrete surface defect classification method based on a residual-self-attention deep fusion network (RSAFuser), aiming to solve the issues of low accuracy and high time consumption in the task. The method uses an "Expert Voting"-like feature extraction approach to identify surface defects of fair-faced concrete, such as Pores, Surface Contamination, Cracks, and Repairs. The residual expert network handles small-area local point defects, such as Pores and Repairs, while the self-attention expert network processes larger-area global linear and block-shaped defects, like cracks and Surface Contamination, extracting and reflecting the overall features of the input image from different perspectives. The fused network not only retains the advantages of the residual network in local information recognition and mitigating network degradation, but also uses the self-attention network for block processing and effectively captures global long-distance information through weight calculation. Experimental results show that compared to ResNet and Swin-Transformer, RSAFuser improves accuracy by 19.19% and 21.51%, respectively. Experimental results validate the effectiveness and accuracy of the proposed method in the fair-faced concrete defect classification task.