<p>Information flow and computational behavior.</p>

Aiming at the imbalance between accuracy and real-time of existing fatigue driving detection models, and the accuracy is lower in low illumination, an LDIE-FDNet (Lightweight Dynamic Image Enhancement-Enabled Real-time Fatigue Driving Detection Network) is designed. Enhance the image by MSR-LIENET (Multi-Scale Retinex-Based Low-Light Image Enhancement Network); Through GSConv_C3k2 module, lightweight design, efficient capture of remote context information, reduction of parameters and calculation, and variable convolution kernel design and feature segmentation and mosaic are adopted to enhance feature extraction ability; Through DHFAR-Net (Dynamic Hierarchical Feature Aggregation and Reconstruction Network), combined with DySample and SDI (Semantic and Detail Infusion), the semantic information and detail information are enhanced. Through multi-level feature fusion, the model can better capture the information of various targets, reduce the situation of missing detection and false detection, thus improving the overall detection effect, and does not need high-resolution boot features as input. It has lower reasoning delay, memory occupation, floating-point operation times, and parameter number; Through the PIoU (Powerfull-IoU) loss function, IoU is calculated pixel by pixel, which can better optimize the positioning of rotating rectangular frames and deal with high aspect ratio targets, reduce the overlap of background areas and improve the detection effect. Finally, a fatigue driving detection model including maximum closing time (MCT) and maximum yawn duration (MYD) is proposed. Experiments show that mAP increases by 0.6% to 99.2, Params reduce by 24%, GFLOPs increase by 14.3% and FPS increases by 23.1% on the YawDD data set. On the DMS data set, mAP increased by 0.7% to 92.9, Params reduce by 24%, GFLOPs increased by 14.3%, and FPS increased by 20.5%. The proposed method enhances both accuracy and efficiency in fatigue detection while effectively balancing precision with real-time performance.针对现有疲劳驾驶检测模型存在的精度与实时性失衡问题，且在低光照环境下检测精度偏低，本文设计了LDIE-FDNet（轻量级动态图像增强型实时疲劳驾驶检测网络，Lightweight Dynamic Image Enhancement-Enabled Real-time Fatigue Driving Detection Network）。首先通过MSR-LIENET（多尺度Retinex低光照图像增强网络，Multi-Scale Retinex-Based Low-Light Image Enhancement Network）对图像进行增强处理；采用GSConv_C3k2模块实现轻量化设计，可高效捕获远程上下文信息，缩减参数量与计算开销，并引入可变卷积核设计、特征分割与特征拼接策略，以强化特征提取能力；通过DHFAR-Net（动态层级特征聚合与重构网络，Dynamic Hierarchical Feature Aggregation and Reconstruction Network），结合DySample与SDI（语义与细节融合模块，Semantic and Detail Infusion），增强语义信息与细节信息的表达；借助多级特征融合，使模型能够更好地捕获多类目标信息，减少漏检与误检情况，进而提升整体检测效果，且无需以高分辨率特征作为输入，具备更低的推理延迟、内存占用、浮点运算量与参数量；采用PIoU（强力交并比，Powerfull-IoU）损失函数，逐像素计算交并比，可更好地优化旋转矩形框的定位效果，适配高长宽比目标，减少背景区域的重叠干扰，进一步提升检测性能。本文最终提出了一种融合最大闭眼时长（maximum closing time, MCT）与最大打哈欠时长（maximum yawn duration, MYD）的疲劳驾驶检测模型。实验结果表明，在YawDD数据集上，mAP提升0.6%至99.2，参数量缩减24%，GFLOPs提升14.3%，FPS提升23.1%；在DMS数据集上，mAP提升0.7%至92.9，参数量缩减24%，GFLOPs提升14.3%，FPS提升20.5%。所提方法在疲劳检测任务中同时提升了检测精度与运行效率，有效平衡了检测精度与实时性需求。