MF-TLID:一种多特征融合输电线覆冰图像去噪方法
An ice-covered transmission line image denoising method fused with multiple features
张宇 1窦银科 2赵亮亮 3焦阳阳 3郭栋梁3
作者信息
- 1. 太原理工大学电气与动力工程学院,太原 030024;太原工业学院自动化系,太原 030008;山西省能源互联网研究院,太原 030032
- 2. 太原理工大学电气与动力工程学院,太原 030024;山西省能源互联网研究院,太原 030032;煤电清洁智能控制教育部重点实验室,太原 030024
- 3. 山西省能源互联网研究院,太原 030032
- 折叠
摘要
针对基于图像对线缆覆冰状态进行监测过程中图像存在噪声的问题,提出了一种多特征融合输电线覆冰图像去噪方法.该方法采用残差注意力融合模块、源特征融合模块和特征增强模块.在残差注意力融合模块中采用级联残差结构和混合注意力模块,既有助于特征信息映射,又能增强特征信息表达;在网络不同特征层中融合源特征,保留图像的低频信息,有利于提升图像的清晰度和真实感;在特征增强模块中同时结合局部和全局特征,通过特征注意力加权学习有效特征向量表示,提高模型去噪能力;创新性地提出像素损失和感知损失的联合损失函数,同时考虑像素级别的误差和感知质量的提升.在输电线覆冰数据集上高斯噪声标准差分别为10～40、20～50和30～60,PSNR 和 SSIM 分别达到了 {31.015 dB、29.262 dB、27.717 dB｝和｛ 0.956、0.943、0.930}.结果表明,该算法的性能优于主流去噪方法,具有更强的抑噪能力和抗干扰性.
Abstract
To address the image noise in monitoring the ice-covered state of transmission lines based on images,this paper proposes an ice-covered transmission line image denoising method fused with multiple features(MF-TLID).The algorithm consists of residual attention fusion module,source feature fusion module and feature enhancement module.The cascaded residual structure and hybrid attention are employed in the residual attention fusion module,which not only contributes to feature information mapping but also enhances the expression of feature information.The source features are fused in different feature layers of the network to retain the low-frequency information of the images,which helps improve the clarity of the image.In the feature enhancement module,both local and global features are combined,and the effective feature vector representation is learned by the feature attention weighting to improve the model removal ability.We propose a joint loss function of Charbonnier loss and Perceptual Loss,taking into account the error of pixel level and the improvement of perceptual quality.On the transmission line icing dataset,the standard deviation of Gaussian noises are between 10-40,20-50 and 30-60,PSNR and SSIM reaches {31.015 dB,29.262 dB,27.717 dB } and {0.956,0.943,0.930} respectively.Our results indicate the proposed method performs better than the mainstream denoising methods,showing stronger noise suppression ability and robustness.
关键词
输电线覆冰/图像去噪/特征融合/注意力机制/联合损失函数Key words
ice-covered transmission line/image denoising/feature fusion/attention mechanism/joint loss function引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据