Abstract: A multi-feature fusion enhancement algorithm based on YOLOv8 is proposed to address the challenges of detecting small bolts in the complex environment and low-resolution conditions at the bottom of a train’s undercarriage. Adaptive contrast stretching is applied during image preprocessing to enhance image quality and highlight bolt details, providing high-quality input for the detection algorithm. The SPD-Conv module is introduced to replace traditional stride convolutions and pooling operations, minimizing fine-grained information loss in small object detection. a BoSTNet architecture is designed to optimize the backbone network and retain small bolt target information effectively. In the Neck layer, a parallel dynamic weighted multi-dimensional fusion attention module is integrated to further suppress noise. In order to accelerate model convergence and improve regression accuracy, the Focaler-MPDIoU function is introduced to optimize the bounding box regression loss so as to efficiently locate the bolt loss function comparison experiments. Experimental results show that, on a custom dataset, the improved YOLOv8 achieves a 3.9, 3.2, and 4.8 percentage points increase in detection accuracy, recall rate, and mAP50, respectively, with values of 95.1%, 94.6%, and 95.0%. This demonstrates the model’s high efficiency in detecting small bolts under complex conditions. Moreover, on the VisDrone-2019 dataset, the improved YOLOv8 outperforms other detection methods, further validating its applicability in complex scenes and small object detection.

Key words: bolt detection, YOLOv8 algorithm, convolutional neural network, hybrid attention mechanism

摘要： 为解决转向架底部复杂环境和低分辨率条件下小目标螺栓检测难度大、精度低的问题，提出一种基于YOLOv8的多特征融合改进算法。在图像预处理阶段，利用自适应式对比度拉伸变换技术增强图像质量，突出螺栓细节特征，为后续目标检测算法提供高质量输入；在特征提取网络中，引入SPD-Conv模块替代传统跨步卷积与池化操作，减少小目标检测中的细粒度信息丢失；在Backbone中，设计一种BoSTNet结构，合理优化主干网络，有效提升对螺栓小目标信息的有效保留；在Neck层中，设计并集成一种并行动态加权多维度融合注意力模块，进一步抑制噪声干扰；为加速模型收敛并提高回归精度，引入Focaler-MPDIoU函数优化边界框回归损失，从而有效地定位螺栓。实验结果表明，在自制数据集上，相比原始模型，改进后的YOLOv8在螺栓检测精度、召回率、mAP50上分别提升3.9、3.2、4.8个百分点，达到95.1%、94.6%和95.0%，体现了该模型在复杂条件下的小目标螺栓的高效性。此外，在VisDrone-2019数据集上，相较于其他检测方法，改进的YOLOv8有更高的检测精度，进一步验证了该模型在复杂场景和小目标检测中的应用价值。

关键词: 螺栓检测, YOLOv8算法, 卷积神经网络, 混合注意力机制