Abstract:

Safety helmets are essential protective gear to ensure physical safety among blasting operators. A Feature Enhancement-YOLOX (FEM-YOLOX) algorithm is proposed to address false target detection and misdetection in helmet detection tasks under low light conditions caused by fuzzy visual information, low brightness, and low contrast. Initially, a Soft Spatial Pyramid Pooling Module (SSPPM) is integrated into the backbone network to minimize feature mapping information loss and to preserve additional context information during downsampling. Second, an Efficient Feature Fusion Module (EFFM) incorporating a lightweight channel attention mechanism is designed to enhance the target area feature learning, improve feature fusion efficiency, and reduce model misdetection. Third, VariFocalLoss is adopted instead of BCEWithlogitsLoss to dynamically adjust the weight of positive and negative samples, making the model focus on fewer positive samples, accelerating the convergence process, and improving the detection accuracy for both target types. Finally, the CIoU is employed as the bounding box regression loss function to enhance the accuracy of the target bounding box prediction. Experimental results demonstrate that the proposed algorithm achieves a mean Average Precision (mAP) improvement of 2.21 percentage points over the baseline and the number of image processed per second increase of 7.67, satisfying the accuracy and speed requirements for real-time safety helmet detection in low light blasting sites.

Key words: safety helmet detection, YOLOX-s algorithm, attention mechanism, bounding box regression loss function, confidence loss function

摘要：

安全帽是保障爆破作业人员人身安全的重要工具。受低照度爆破现场安全帽检测任务中目标视觉信息模糊、图像亮度低及对比度低的影响, 在目标检测过程中存在目标漏检、误检等问题。基于YOLOX提出了特征增强的安全帽检测算法FEM-YOLOX。首先, 在主干网络使用软池化构建软空间金字塔池化模块(SSPPM), 减少了特征映射中的信息弥散, 并在下采样映射中保留了更多上下文信息; 其次, 设计基于高效通道注意力(ECA)机制的高效特征融合模块(EFFM), 加强了模型对目标区域特征的学习, 提高了特征融合的效率, 减少了模型误检情况的出现; 再次, 采用VariFocalLoss替代BCEWithlogitsLoss, 动态调整正负样本的权重, 使得模型关注数量较少的正样本, 加速了模型的收敛过程, 提升了两类目标的检测精度; 最后, 采用CIoU作为边框回归损失函数, 提高了模型定位目标预测框的精度。实验结果表明, 所提算法的均值平均精度(mAP)相较于基线算法提升了2.21百分点, 每秒处理的图像数量提升了7.67, 满足了低照度爆破现场安全帽实时检测的精度和速度需要。

关键词: 安全帽检测, YOLOX-s算法, 注意力机制, 边框回归损失函数, 置信度损失函数