Learning with noisy labels for robust fatigue detection

Mei Wang, Ruimin Hu, Xiaojie Zhu, Dongliang Zhu, Xiaochen Wang

Abstract: Fatigue is a significant safety concern across various domains, and accurate detection is vital. However, the commonly employed fine-grained labels (seconds-based) frequently inherit coarse-grained labels (minutes-based or more), inevitably introducing noise due to fatigue states’ dynamic and time-varying nature. Compared to noise in existing image tasks, fatigue noise is complex and diverse, with relatively ambiguous category boundaries. To address this issue, we propose a novel class-level fatigue noise-tolerant supervised comparative learning (cFNSCL) method that explores the data structure within each class, extracts trustworthy samples, and encourages the learning of distinguishable representations against noise. Specifically, supervised contrastive learning (SCL) is introduced to deal with complex and variable noise, and a dynamic noise-tolerant contrastive loss (DNCL) that incorporates a novel class-level confidence assessment mechanism (CCAM) is designed. CCAM selects high-confidence samples for DNCL learning, significantly alleviating SCL’s sensitivity to noise and enhancing the model’s tolerance to noise. Additionally, a novel class-level trustworthy sample extraction (cTSE) method from the perspective of inherited label categories to improve the model’s representative ability is proposed. Experimental results demonstrate the effectiveness of cFNSCL on both synthetic and real-world noisy datasets over some state-of-the-art methods. Specifically, it improves the accuracy by an average of 6.11% and 6.95% on the two real-world datasets, respectively.