基于多类别Focal Loss损失函数的变电站场景图像语义分割研究

Research on Semantic Segmentation of Transformer Substation Image Based on Multi-category Focal Loss Function

图像语义分割技术在变电站智能机器人抢险作业以及日常巡检中起重要作用。针对变电站场景中的类别不平衡问题,提出了一种用于训练多类别图像语义分割模型的多类别Focal Loss损失函数。多类别Focal Loss损失函数可以动态调整各类别的权值,对出现频率小的类别更加友好。基于深度学习和变电站场景图像,分别使用多类别Focal Loss损失函数和交叉熵损失函数训练基于FCN、SegNet和DeepLabV3网络的图像语义分割模型进行对比实验,并通过平均交并比和像素准确率等评价指标进行模型评价。实验结果表明基于多类别Focal Loss损失函数的图像语义分割模型具有良好效果,有助于缓解类别不平衡现象。

Image semantic segmentation technology plays an important role in the rescue operations and daily inspections of intelligent robots in substations.In order to solve the problem of category imbalance in substation scenes,a multi-category Focal Loss function for training the semantic segmentation model of multi-category image semantic segmentation was proposed.The multi-category Focal Loss function can dynamically adjust the weight of each category,which is more friendly to categories that appear less frequently.Based on deep learning and scene images of substations,we used multi-category Focal Loss function and cross entropy loss function to train image semantic segmentation models for comparative experiments based on the FCN,SegNet and DeepLabV3 networks,and the model was evaluated by metrics such as average intersection ratio and pixel accuracy.The experimental results show that the image semantic segmentation model based on the multi-category Focal Loss function has a good effect and helps to alleviate the category imbalance.