【摘要】 针对变电站设备腐蚀不易检测且难以定量化的问题,提出了一种改进的SOLOv2实例分割模型,并构建了配套的腐蚀程度定量化分级方法。首先,将模型的主干网络由ResNte更换为ResNeXt,以减少模型计算量,增强模型的适应性;其次,颈部部分使用PaFPN用以融合特征,增强特征融合能力;最后,在掩码特征分支中嵌入SimAM注意力机制,以提升模型分割精度。基于实地采集的变电站设备数据集对所提网络进行训练与测试。结果表明:改进后的模型性能显著优于原始SOLOv2,其平均精度均值(mAP)提升了6.4%;与当前主流实例分割模型相比,该模型在检测精度与推理速度上均展现出竞争优势,充分验证了改进策略的有效性。基于高精度的分割结果,成功实现了变电站设备腐蚀等级的自动化定量化分级。更多还原

【Abstract】 To address the challenges of detecting and quantifying corrosion in substation equipment, an improved SOLOv2 instance segmentation model was proposed, along with a corresponding method for the quantitative grading of corrosion severity. First, the backbone network of the model was replaced with Res Ne Xt to reduce computational complexity and enhance model adaptability. Second, the path aggregation feature pyramid network(Pa FPN) was introduced in the neck module to improve feature fusion capability. Finally, the Sim AM attention mechanism was embedded into the mask feature branch to further improve segmentation accuracy. The proposed network was trained and evaluated on a dataset of substation equipment images collected in the field. The results show that the improved model significantly outperformed the original SOLOv2, achieving a 6.4% increase in mean average precision(m AP). Compared with current mainstream instance segmentation models, the proposed model exhibited competitive advantages in both detection accuracy and inference speed, fully validating the effectiveness of the improvement strategies. Furthermore, based on the high-precision segmentation results, the automatic quantitative grading of corrosion levels in substation equipment was successfully achieved.更多还原