深立井井壁图像的卷积神经网络去噪方法
|
摘要
为了实现井壁缺陷的自动检测,提出去除井壁图像噪声的卷积神经网络(CNN)模型(ELU-CNN)。该模型为深28层的全卷积网络模型,由5个特征提取模块(FEM)和跳跃连接组成;跳跃连接将第一卷积层的输出特征与每一个FEM的输出特征串联融合,保证图像特征的充分提取;使用残差学习来缓解梯度消失并提高收敛速度,保证训练后的去噪模型学习到的非线性映射是图像噪声;选用ELU作为激活函数,它具有软饱和性且输出均值接近于零,能增强模型对输入噪声的鲁棒性并加速模型收敛。在标准测试集BSD68、set12及实际井壁图像上,验证ELU-CNN模型的去噪性能并和先进方法作比较,实验结果表明:与FFDNet模型相比,ELU-CNN模型的平均峰值信噪比,在含噪声浓度σ为(15,25,35,50,75)的BSD68、set12测试集上分别提高了(0.17,0.11,0.08,0.05,0.03) dB、(0.18,0.16,0.08,0.06,0.07) dB。在去除井壁图像盲噪声时,ELU-CNN模型能更好地保留缺陷的纹理信息。
An ELU-CNN image denoising model based on convolution neural network(CNN) is proposed to realize automatic detection of the defects on the shaft lining. It is a 28-layer full-convolution denoising network model and consists of five feature extraction modules(FEMs) and skip connections. The skip connections combine the output features of the first convolution layer with the output features of each FEM to ensure the full extraction of image features. The residual learning is used in the ELU-CNN model to relieve the gradient disappearance problem, to improve the convergence speed, and to ensure that the learned nonlinear mapping from a noisy image by the denoising model after training is a noise image. ELU is used as an activation function. It has soft saturation performance, and the average of its outputs is close to zero. These properties can accelerate the convergence of the model and enhance the robustness of the model to the input noises. The denoising effect of ELU-CNN is verified on the standard test sets BSD68 and set12 as well as actual shaft wall images, and is compared with those of some advanced methods. The experimental results on BSD68 and set12 with noise concentration σ=(15, 25, 35, 50, 75) and comparisons with FFDNet model show that the average of peak signal to noise ratios of ELU-CNN increases(0.17, 0.11, 0.08, 0.05, 0.03) dB and(0.18, 0.16, 0.08, 0.06, 0.07) dB, respectively. ELU-CNN can better preserve the texture information of fractures when removing blind noise from shaft wall images.
An ELU-CNN image denoising model based on convolution neural network(CNN) is proposed to realize automatic detection of the defects on the shaft lining. It is a 28-layer full-convolution denoising network model and consists of five feature extraction modules(FEMs) and skip connections. The skip connections combine the output features of the first convolution layer with the output features of each FEM to ensure the full extraction of image features. The residual learning is used in the ELU-CNN model to relieve the gradient disappearance problem, to improve the convergence speed, and to ensure that the learned nonlinear mapping from a noisy image by the denoising model after training is a noise image. ELU is used as an activation function. It has soft saturation performance, and the average of its outputs is close to zero. These properties can accelerate the convergence of the model and enhance the robustness of the model to the input noises. The denoising effect of ELU-CNN is verified on the standard test sets BSD68 and set12 as well as actual shaft wall images, and is compared with those of some advanced methods. The experimental results on BSD68 and set12 with noise concentration σ=(15, 25, 35, 50, 75) and comparisons with FFDNet model show that the average of peak signal to noise ratios of ELU-CNN increases(0.17, 0.11, 0.08, 0.05, 0.03) dB and(0.18, 0.16, 0.08, 0.06, 0.07) dB, respectively. ELU-CNN can better preserve the texture information of fractures when removing blind noise from shaft wall images.
引文
[1] 何满潮,袁和生,靖洪文.中国煤矿锚杆支护理论与实践 [M].北京:科学出版社,2004:5-32.
[2] DIERING D H.Ultra-deep level mining:future requirements [J].Journal of South African Institute of Mining and Metallurgy,1997,97(6):249-255.
[3] 阮秋琦.数字图像处理学 [M].北京:电子工业出版社,2013:1-30.
[4] CHATTERJEE P,MILANFAR P.Is denoising dead?[J].IEEE Transactions on Image Processing,2010,19(4):895-911.
[5] DABOV K,FOI A,KATKOVNIK V,et al.Image denoising by sparse 3-D transform-domain collaborative filtering [J].IEEE Transactions on Image Processing,2007,16(8):2080-2095.
[6] GU Shuhang,ZHANG Lei,ZUO Wangmeng,et al.Weighted nuclear norm minimization with application to image denoising [C]//Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2014:2862-2869.
[7] SCHMIDT U,ROTH S.Shrinkage fields for effective image restoration [C]//Proceedings of the 27th IEEE Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2014:2774-2781.
[8] CHEN Y J,POCK T.Trainable nonlinear reaction diffusion:a flexible framework for fast and effective image restoration [J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(6):1256 -1272.
[9] ZHANG Kai,ZUO Wangmeng,CHEN Yunjin,et al.Beyond a Gaussian denoiser:residual learning of deep CNN for image denoising [J].IEEE Transactions on Image Processing,2017,26(7):3142-3155.
[10] ZHANG Kai,ZUO Wangmeng,ZHANG Lei.FFDNet:toward a fast and flexible solution for CNN-based image denoising [J].IEEE Transactions on Image Processing,2018,27(9):4608-4622.
[11] ISOGAWA K,IDA T,SHIODERA T,et al.Deep shrinkage convolutional neural network for adaptive noise reduction [J].IEEE Signal Processing Letters,2017,25(2):224-228.
[12] HUANG Gao,LIU Zhuang,LAURENS V D M.Densely connected convolutional networks [C]//Proceedings of the 30th IEEE Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2018:2261-2269.
[13] JAIN V,SEUNG S.Natural image denoising with convolutional networks [C]//Proceedings of the Neural Information Processing Systems.New York,USA:Curran Associates Inc.,2009:769-776.
[14] BURGER H C,SCHULER C J,HARMELING S.Image denoising:can plain neural networks compete with BM3D?[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2012:2392-2399.
[15] NAIR V,HINTON G E.Rectified linear units improve restricted Boltzmann machines [C]//Proceedings of the 27th International Conference on Machine Learning.New York,USA:ICML,2010:807-814.
[16] CLEVERT D A,UNTERTHINER T,HOCHREITER S.Fast and accurate deep network learning by exponential linear units (ELUs) [EB/OL].(2016-02-22) [2018-11-17].https://arxiv.org/abs/1511.07289 v5.
[17] BURGER H C,SCHULER C,HARMELING S.Learning how to combine internal and external denoising methods [C]//Proceedings of the 35th German Conference on Pattern Recognition.Berlin,Germany:Springer Verlag,2013:121-130.
[2] DIERING D H.Ultra-deep level mining:future requirements [J].Journal of South African Institute of Mining and Metallurgy,1997,97(6):249-255.
[3] 阮秋琦.数字图像处理学 [M].北京:电子工业出版社,2013:1-30.
[4] CHATTERJEE P,MILANFAR P.Is denoising dead?[J].IEEE Transactions on Image Processing,2010,19(4):895-911.
[5] DABOV K,FOI A,KATKOVNIK V,et al.Image denoising by sparse 3-D transform-domain collaborative filtering [J].IEEE Transactions on Image Processing,2007,16(8):2080-2095.
[6] GU Shuhang,ZHANG Lei,ZUO Wangmeng,et al.Weighted nuclear norm minimization with application to image denoising [C]//Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2014:2862-2869.
[7] SCHMIDT U,ROTH S.Shrinkage fields for effective image restoration [C]//Proceedings of the 27th IEEE Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2014:2774-2781.
[8] CHEN Y J,POCK T.Trainable nonlinear reaction diffusion:a flexible framework for fast and effective image restoration [J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(6):1256 -1272.
[9] ZHANG Kai,ZUO Wangmeng,CHEN Yunjin,et al.Beyond a Gaussian denoiser:residual learning of deep CNN for image denoising [J].IEEE Transactions on Image Processing,2017,26(7):3142-3155.
[10] ZHANG Kai,ZUO Wangmeng,ZHANG Lei.FFDNet:toward a fast and flexible solution for CNN-based image denoising [J].IEEE Transactions on Image Processing,2018,27(9):4608-4622.
[11] ISOGAWA K,IDA T,SHIODERA T,et al.Deep shrinkage convolutional neural network for adaptive noise reduction [J].IEEE Signal Processing Letters,2017,25(2):224-228.
[12] HUANG Gao,LIU Zhuang,LAURENS V D M.Densely connected convolutional networks [C]//Proceedings of the 30th IEEE Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2018:2261-2269.
[13] JAIN V,SEUNG S.Natural image denoising with convolutional networks [C]//Proceedings of the Neural Information Processing Systems.New York,USA:Curran Associates Inc.,2009:769-776.
[14] BURGER H C,SCHULER C J,HARMELING S.Image denoising:can plain neural networks compete with BM3D?[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.Piscataway,NJ,USA:IEEE,2012:2392-2399.
[15] NAIR V,HINTON G E.Rectified linear units improve restricted Boltzmann machines [C]//Proceedings of the 27th International Conference on Machine Learning.New York,USA:ICML,2010:807-814.
[16] CLEVERT D A,UNTERTHINER T,HOCHREITER S.Fast and accurate deep network learning by exponential linear units (ELUs) [EB/OL].(2016-02-22) [2018-11-17].https://arxiv.org/abs/1511.07289 v5.
[17] BURGER H C,SCHULER C,HARMELING S.Learning how to combine internal and external denoising methods [C]//Proceedings of the 35th German Conference on Pattern Recognition.Berlin,Germany:Springer Verlag,2013:121-130.