Multi-label Classification of Road Cracks Based on Semantic-Aware Mixup

Chengyang Xu, Chunpeng Zhou, Zhi Yu

Abstract: In the realm of road crack detection scenarios, Mixup, a widely used data augmentation technique, exhibits constrained effectiveness in multi-label classification, unlike its performance on general datasets. This paper investigates the reasons behind this limitation, identifying two main factors: (1) the mutually exclusive nature of road crack labels and (2) the potential generation of incorrect labels during image mixing. To address these issues, we propose an approach called Semantic-Aware Mixup (SA-Mixup), which focuses on enhancing crack samples specifically. Experimental evaluations on three distinct road crack datasets demonstrate the effectiveness and simplicity of our proposed method. The results highlight significant improvements in model performance, offering a promising solution for accurate and efficient multi-label classification in road crack detection tasks.