基于改进Mask R-CNN的变电设备红外图像实例分割算法

Instance Segmentation Algorithm for Substation Equipment Infrared Images Based on Improved Mask R-CNN

红外图像中变电设备的分割精度直接影响着故障诊断的结果,针对复杂红外背景下变电设备边缘分割不精细、分割精度低的问题,提出了一种基于改进Mask R-CNN模型的变电设备红外图像分割方法。首先将ResNet特征提取网络中部分残差模块的标准卷积替换为可变形卷积,然后对空间注意力机制模块和通道注意力机制模块并行连接,并在这两个模块中加入可变形卷积,最后改进Mask R-CNN掩膜分支的损失函数,对目标边缘分割的精细度进一步优化。该方法能够有效提高模型对红外图像中变电设备几何特征多样性的适应能力,并减轻模型对背景等干扰特征的关注。在变电设备红外图像数据集上进行实验,结果表明,相比于Mask R-CNN基准模型,该方法的AP_(50:95)、AP_(50)和AP_(75)提高了3.5%、1.0%、4.2%,表明该方法能够显著提高红外图像中变电设备实例分割的准确率,有效解决边缘分割不精细的问题。

The accuracy of substation equipment extraction in Infrared images directly affects the results of fault diagnosis.To solve the problem that the edge segmentation of substation equipment in the complex infrared background is imprecise and the segmentation accuracy is low,an infrared image segmentation method for substation equipment based on improved Mask R-CNN model was proposed.Firstly,the standard convolution of some residual modules in the ResNet feature extraction network was replaced with a deformable convolution.Then the spatial attention mechanism module and the channel attention mechanism module were connected in parallel,and deformable convolution was added to both modules.Finally,the loss function of the Mask R-CNN mask branch was improved to further optimize the fineness of target edge segmentation.The method can effectively improve the model’[KG-*3]s ability to adapt to the diversity of geometric features of substation equipment in infrared images and alleviate the model’[KG-*3]s focus on interference features such as the background.Experimental validations were carried out on the dataset of infrared images of substation equipment and the results show that AP_(50:95),AP_(50)and AP_(75)improved by the method in this paper are respectively 3.5%,1.0%,and 4.2%higher than that improved by the Mask R-CNN benchmark model.It is shown that the method can significantly improve the accuracy of the segmentation of substation equipment in infrared images and effectively solve the problem of imprecise edge segmentation.