This paper proposes a lightweight YOLO object detection algorithm based on bidirectional multi‐scale feature enhancement. The problem is that the original YOLOv5 algorithm does not make full use of the relationship between the feature layers, resulting in the loss of target semantic information and a large number of parameters. First, a bidirectional multi‐scale feature‐enhanced weighted fusion backbone network is constructed to extract target features repeatedly. It enhances the fusion ability of shallow detail features and high‐level semantic information to capture richer multi‐scale semantic information. Second, the NCA attention module is built and integrated into the feature fusion network to enhance the critical characteristics of the target region. Finally, the Ghost module is used instead of the convolutional blocks in the original network to lighten the model while reducing the network complexity and training difficulty. Experimental results show that the improved YOLOv5 algorithm achieves 78.8% mAP@0.5 for the PASCAL VOC2012 dataset, which is 1.5% higher than the original algorithm, at 62.5 FPS. The number of parameters is also reduced by 43.6%. The mAP@0.5 on the self‐made metal foreign object dataset reached 98.4%, at 58.8 FPS, which can meet the requirements of end‐device deployment and real‐time detection.

中文翻译：

---

一种基于双向多尺度特征增强的轻量级YOLO目标检测算法

本文提出一种基于双向多尺度特征增强的轻量级YOLO目标检测算法。问题在于原始YOLOv5算法没有充分利用特征层之间的关系，导致目标语义信息丢失和大量参数。首先，构建双向多尺度特征增强加权融合主干网络来重复提取目标特征。它增强了浅层细节特征和高层语义信息的融合能力，以捕获更丰富的多尺度语义信息。其次，构建NCA注意力模块并将其集成到特征融合网络中，以增强目标区域的关键特征。最后，使用Ghost模块代替原网络中的卷积块，减轻模型重量，同时降低网络复杂度和训练难度。实验结果表明，改进后的YOLOv5算法在PASCAL VOC2012数据集上达到了78.8% mAP@0.5，比原算法提高了1.5%，帧速率为62.5 FPS。参数数量也减少了43.6%。自制金属异物数据集上的mAP@0.5达到98.4%，帧速率为58.8 FPS，可以满足终端设备部署和实时检测的要求。