摘要
针对红外与可见光异源图像特征不同的模态、语义与空间特性,提出一种多尺度异源图像协同的电力设备语义分割方法。首先利用双路编码网络对红外图像和可见光图像进行特征提取,并以超像素为单位对红外图像高层特征进行增强,在抑制背景噪声的同时,对高级特征进行进一步的融合,提取更多的互补语义信息。同时,在高级语义信息引导下,在通道维度对中低级特征进行融合增强。最后,通过多层次解码,逐步结合多尺度特征,并以多标签损失的方式,实现更精准的语义分割。通过自建电力设备语义分割数据集进行对比与消融实验,实验结果证明提出融合策略的有效性与语义分割网络的优秀性能。
In view of the different modality,semantic and spatial characteristics of infrared and visible heterogeneous image features,we propose a semantic segmentation method for power equipment based on differentiated fusion of multi-scale heterogeneous features.First,a dual-stream encoding network is used to extract thermal infrared(T)and visible images(red-green-blue,RGB)features.And the high-level features of infrared images are enhanced in units of super-pixels.While suppressing background noise,the high-level features are further integrated to extract more complementary semantic information.Meanwhile,guided by high-level semantic information,the mid-and low-level features are fused and enhanced in the channel dimension.Finally,through multi-level decoding,multi-scale features are gradually combined and multi-label loss function is used to achieve more accurate semantic segmentation.Extensive comparison and ablation experiments were conducted on the self-built semantic segmentation dataset of electric power equipment.The experimental results prove effectiveness of the fusion strategy and the excellent performance of the proposed semantic segmentation network.
作者
李庆武
方保民
孟凡领
赵书楷
贺卫刚
惠远鑫
LI Qingwu;FANG Baomin;MENG Fanling;ZHAO Shukai;HE Weigang;HUI Yuanxin(College of Information Science and Engineering,Hohai University,Changzhou 213251,China;State Grid Qinghai Electric Power Company,Xining 630100,China;State Grid Haidong Power Supply Company,Haidong 630200,China)
出处
《应用科技》
CAS
2024年第5期58-65,共8页
Applied Science and Technology
基金
国网青海省电力公司科技项目(SGQHHD00YXJS2310532)。
关键词
多尺度异源图像特征
特征融合
红外图像
可见光图像
电力设备
语义分割
超像素分割
多标签预测
multi-scale heterogeneous image features
feature fusion
infrared images
visible images
power equipment
semantic segmentation
hyperpixel segmentation
multi-label prediction
