【摘要】 巡检机器人拍摄的变电设备红外图像往往具有背景复杂、目标重叠、截断以及远处目标尺度小等特点，识别难度较大，因此本文提出一种改进YOLOv8的变电设备红外图像目标识别模型。首先，引入Soft＿NMS减少重叠目标丢失问题；其次，在颈部网络添加多维协作注意力机制MCA,使网络聚焦于相关的特征区域，并采用Inner-IoU和Focal loss混合损失函数，增强模型对小尺度目标的泛化能力以及对高质量锚框的关注度；最后，采用GhostNetV2模块对模型进行轻量化设计。通过实验表明，本模型较YOLOv8n基准模型，平均精度均值mAP@0.5提升4.9%、模型参量减少36.2%、检测速度达到160.1FPS,有效提高了模型的识别能力和轻量化水平，为后续的变电设备故障诊断提供基础。更多还原

【Abstract】 Infrared images of substation equipment taken by the inspection robot often have the characteristics of complex background, target overlap, truncation and small scale of distant targets, posing significant challenges to recognition.Therefore, an improved infrared image target recognition model for substation equipment with YOLOv8 so is proposed in this paper.Firstly, Soft＿NMS is introduced to reduce the loss of overlapping targets.Secondly, a multi-dimensional cooperative attention mechanism MCA is added to the neck network to focus on the relevant feature regions, and the mixed loss function of Inner-IoU and Focal loss is adopted to enhance the model′s generalization ability to small-scale targets and the attention to high-quality anchor frames.Finally, the GhostNetV2 module is employed for lightweight design of the model.Experimental results show that compared with the YOLOv8n benchmark model, the mAP@0.5 of this model is increased by 4.9 %,the model parameters are reduced by 36.2 %,and the detection speed reaches 160.1 FPS,which effectively improves the recognition ability and lightweight level of the model, laying a foundation for subsequent fault diagnosis of substation equipment.更多还原