首页|基于注意力引导多尺度降噪卷积神经网络的钢轨表面缺陷图像降噪

基于注意力引导多尺度降噪卷积神经网络的钢轨表面缺陷图像降噪

扫码查看
原文链接
  • 万方数据
  • 维普
针对钢轨表面缺陷图像降噪依赖人工设置滤波参数和缺陷边缘模糊的问题,提出基于注意力引导多尺度降噪卷积神经网络的钢轨表面缺陷图像降噪方法.首先采用深层网络中的多尺度卷积自动提取含噪图像的特征,使其不依赖于人工设置滤波参数,并克服单尺度卷积特征不够精细导致缺陷边缘模糊的问题;其次利用跳跃连接融合网络深层特征和浅层特征,强化浅层特征影响,克服因网络加深导致浅层特征被忽略的问题,使特征更充分;然后利用注意力机制调节特征在空间不同位置的权重,筛选出能表征噪声的特征,获得噪声信息;最后通过重建模块去除含噪图像中的噪声,实现端到端的降噪.试验结果从定性和定量角度证明所提方法不仅降噪效果更好,且更有效地保留了缺陷边缘信息,为缺陷精确分割提供条件.
Noise Reduction of Rail Surface Defect Images Based on Attention-guided Poly-scale Denoising Convolutional Neural Networks
In order to solve the problem of the dependence of noise reduction of rail surface defect images on manual set-ting of filtering parameters and blurring of defect edges,an attention-guided poly-scale noise reduction convolutional neu-ral network-based noise reduction method was proposed for rail surface defect images.Firstly,the poly-scale convolution in the deep network was used to automatically extract the features of noise-containing images,to avoid relying on manu-ally set filtering parameters and overcome the problem of blurred defect edges caused by insufficient refinement of single-scale convolutional features.Secondly,the deep and shallow features of the network were fused using jumping connec-tions to strengthen the influence of the shallow features and overcome the problem of the shallow features being ignored in the deep layer due to the deeper network,to deliver more adequate features.Thirdly,attention mechanism was used to adjust the weights of features at different locations in space to filter out features that can characterize noise,and obtain noise information.Finally,the noise information in noisy images was removed by the reconstruction module to achieve end-to-end noise reduction.The experimental results demonstrate qualitatively and quantitatively that the proposed method is more effective both in noise reduction and in retaining defect edge information,providing conditions for accu-rate defect segmentation.

rail surface defectimage denoisingconvolutional neural networkpoly-scale feature

陈仁祥、潘升、杨黎霞、王建西、夏天

展开 >

重庆交通大学交通工程应用机器人重庆市工程实验室,重庆 400074

重庆科技大学工商管理学院,重庆 401331

石家庄铁道大学道路与铁道工程安全保障教育部重点实验室,河北石家庄 050043

钢轨表面缺陷 图像降噪 卷积神经网络 多尺度特征

国家自然科学基金道路与铁道工程安全保障省部共建教育部重点实验室开放基金重庆交通大学市级研究生联合培养基地项目重庆市教委科学技术研究计划重庆交通大学市级专业学位研究生教学案例库项目重庆交通大学研究生科研创新项目

51975079STDTKF202204JDLHPYJD2021007KJZD-M202200701JDALK20220072022S0045

2024

10.3969/j.issn.1001-8360.2024.05.015

铁道学报
中国铁道学会

铁道学报

CSTPCD北大核心
影响因子：0.9
ISSN：1001-8360
年,卷(期)：2024.46(5)
  • 7