采用雙通道卷積神經(jīng)網(wǎng)絡(luò)構(gòu)建的隨機(jī)脈沖噪聲深度降噪模型

徐少平; 林珍玉; 崔燕; 劉蕊蕊; 楊曉輝

doi:10.11999/JEIT190796

采用雙通道卷積神經(jīng)網(wǎng)絡(luò)構(gòu)建的隨機(jī)脈沖噪聲深度降噪模型

doi: 10.11999/JEIT190796

南昌大學(xué)信息工程學(xué)院南昌 330031

基金項(xiàng)目: 國家自然科學(xué)基金(61662044, 61163023)，江西省自然科學(xué)基金(20171BAB202017)

詳細(xì)信息

作者簡介:
徐少平：男，1976年生，博士，教授，博士生導(dǎo)師，主要研究方向?yàn)閳D形圖像處理技術(shù)、機(jī)器視覺、虛擬手術(shù)仿真

林珍玉：女，1996年生，碩士生，研究方向?yàn)閳D形圖像處理技術(shù)、機(jī)器視覺

崔燕：女，1996年生，碩士生，研究方向?yàn)閳D形圖像處理技術(shù)、機(jī)器視覺

劉蕊蕊：女，1995年生，碩士生，研究方向?yàn)閳D形圖像處理技術(shù)、機(jī)器視覺

楊曉輝：男，1978年生，博士，副教授，主要研究方向?yàn)楣收显\斷及圖像處理

通訊作者:
徐少平　xushaoping@ncu.edu.cn

中圖分類號(hào): TN911.73; TP391
計(jì)量
- 文章訪問數(shù): 2164
- HTML全文瀏覽量: 801
- PDF下載量: 73
- 被引次數(shù): 0
出版歷程
- 收稿日期: 2019-10-16
- 修回日期: 2020-07-20
- 網(wǎng)絡(luò)出版日期: 2020-07-30
- 刊出日期: 2020-10-13

A Dual-Channel Deep Convolutional Neural Network Model for Random-Valued Impulse Noise Removal

School of Information Engineering, Nanchang University, Nanchang 330031, China

Funds: The National Natural Science Foundation of China (61662044, 61163023), The Natural Science Foundation of Jiangxi Province (20171BAB202017)

摘要

摘要: 為提高對(duì)隨機(jī)脈沖噪聲(RVIN)圖像的降噪效果，該文提出一種被稱為雙通道降噪卷積神經(jīng)網(wǎng)絡(luò)(D-DnCNN)的RVIN深度降噪模型。首先，提取多個(gè)不同階對(duì)數(shù)差值排序(ROLD)統(tǒng)計(jì)值及1個(gè)邊緣特征統(tǒng)計(jì)值構(gòu)成描述圖塊中心像素點(diǎn)是否為RVIN噪聲的噪聲感知特征矢量。其次，利用預(yù)先訓(xùn)練好的深度置信網(wǎng)絡(luò)(DBN)預(yù)測(cè)模型實(shí)現(xiàn)特征矢量到噪聲標(biāo)簽的映射，完成對(duì)噪聲圖像中噪聲點(diǎn)的檢測(cè)。再次，在噪聲檢測(cè)標(biāo)簽的指示下采用Delaunay三角剖分插值算法快速修復(fù)噪聲像素點(diǎn)從而獲得初步復(fù)原圖像。最后，將初步復(fù)原圖像作為參考圖像與噪聲圖像聯(lián)接(concatenate)后輸入D-DnCNN模型后獲得殘差圖像，將參考圖像減去殘差圖像即可獲得降噪后圖像。實(shí)驗(yàn)數(shù)據(jù)表明：D-DnCNN模型在各個(gè)噪聲比例下的降噪效果均顯著超過了現(xiàn)有的經(jīng)典開關(guān)型RVIN降噪算法，與普通的單通道RVIN深度降噪模型相比也有較大幅度提升。
- 圖像處理 /
- 隨機(jī)脈沖噪聲 /
- 雙通道降噪卷積神經(jīng)網(wǎng)絡(luò) /
- 參考圖像 /
- 噪聲感知特征 /
- 噪聲檢測(cè) /
- 插值
Abstract: A Dual-channel Denoising Convolutional Neural Network (D-DnCNN) model for the removal of Random-Valued Impulse Noise (RVIN) is proposed. To obtain the reference image quickly, several Rank-Ordered Logarithmic absolute Difference (ROLD) statistics and one edge feature statistic are first extracted from a local window to construct a RVIN-aware feature vector which can describe the central pixel of the patch is RVIN or not. Next, a noise detector based on Deep Belief Network (DBN) is trained to map the extracted feature vectors to their corresponding noise labels to detect all noise-like pixels in the observed image. Then, under the guidance of noise labels, the Delaunay triangulation-based interpolation algorithm is exploited to restore all detected noise-like pixels quickly and generate a preliminary restored image used as reference image. Finally, the reference image and the noisy image are simultaneously fed into the D-DnCNN model to output its corresponding residual image, and the final restored image can be obtained by subtracting the residual image from the noisy image. Extensive experimental results show that, the denoising effect of the proposed D-DnCNN denoising model outperforms the existing state-of-art switching ones across a range of noise ratios, and it also works better than the ordinary single-channel DnCNN model.
- Image processing /
- Random-Valued Impulse Noise(RVIN) /
- Dual-channel Denoising Convolutional Neural Network(D-DnCNN) /
- Reference image /
- Noise-aware feature /
- Noise detection /
- Interpolation

HTML全文

圖 1 基于DBN的噪聲標(biāo)簽矩陣生成流程

下載: 全尺寸圖片幻燈片

圖 2 利用Delaunay三角剖分插值算法對(duì)Lena噪聲圖像復(fù)原效果

下載: 全尺寸圖片幻燈片

圖 3 帶輔助通道的CNN深度卷積神經(jīng)網(wǎng)絡(luò)的RVIN降噪模型框架

下載: 全尺寸圖片幻燈片

圖 4 各算法對(duì)Lena圖像降噪的效果對(duì)比

下載: 全尺寸圖片幻燈片

表 1 DBN網(wǎng)絡(luò)在Set12測(cè)試集圖像上的預(yù)測(cè)準(zhǔn)確性

圖像	20%噪聲			40%噪聲			60%噪聲			檢測(cè)正確率均值
圖像	False	Miss	Accuracy	False	Miss	Accuracy	False	Miss	Accuracy	檢測(cè)正確率均值
Cameraman	838	2257	0.9528	1914	3952	0.9105	3863	4062	0.8791	0.9141
House	209	1896	0.9679	911	3665	0.9302	2430	4123	0.9000	0.9327
Peppers	400	2524	0.9554	1254	4402	0.9137	3462	4489	0.8787	0.9159
Starfish	536	3217	0.9427	1594	5753	0.8879	5558	4647	0.8443	0.8916
Monarch	489	2776	0.9502	1773	4788	0.8999	4291	4313	0.8687	0.9063
Airplane	1108	2516	0.9447	1979	4514	0.9009	4286	4203	0.8705	0.9054
Parrot	588	2723	0.9495	1877	4465	0.9032	4374	4204	0.8691	0.9073
Lena	755	8303	0.9654	2342	15574	0.9317	9976	17336	0.8958	0.9310
Barbara	2219	12393	0.9443	8329	22147	0.8837	25515	18555	0.8319	0.8866
Boat	1758	10620	0.9564	5318	19001	0.9072	16137	18645	0.8673	0.9103
Man	1714	9717	0.9564	3976	17712	0.9173	13760	18459	0.8771	0.9169
Couple	2027	11049	0.9501	5553	19695	0.9037	16993	19032	0.8626	0.9055

下載: 導(dǎo)出CSV

表 2 不同噪聲比例下各個(gè)降噪算法在BSD68測(cè)試圖像集上所獲得的PSNR均值 (dB)

算法	噪聲比例(%)
算法	10	20	30	40	50	60
ROLD-EPR	30.24	28.26	26.97	25.96	25.04	23.98
ASWM	28.90	27.99	27.01	25.82	23.84	21.05
ROR-NLM	27.29	26.67	25.88	24.69	22.73	20.14
WCSR	30.11	27.93	26.55	25.51	24.52	23.49
ALOHA	31.75	29.04	25.13	23.74	21.81	18.79
WIN5-RB	34.67	31.46	29.02	27.11	25.46	23.68
RED-Net	33.11	30.68	28.87	27.29	25.81	24.37
LSM-NLR	28.86	26.85	25.59	24.63	23.76	22.86
S-DnCNN	35.76	32.41	30.10	27.79	26.15	24.20
本文D-DnCNN	35.71	32.72	30.56	28.62	26.76	25.31

下載: 導(dǎo)出CSV

表 3 D-DnCNN與S-DnCNN算法在真實(shí)噪聲圖像集上降噪效果PSNR對(duì)比(dB)

對(duì)比算法	圖像編號(hào)										均值
對(duì)比算法	1	2	3	4	5	6	7	8	9	10	均值
S-DnCNN	46.85	43.79	52.98	49.64	47.54	43.52	52.47	43.58	42.24	40.66	46.32
本文D-DnCNN	47.45	44.56	54.20	50.32	48.27	44.10	53.81	45.26	43.17	43.17	47.43

下載: 導(dǎo)出CSV

表 4 各算法執(zhí)行時(shí)間的比較(s)

算法	執(zhí)行時(shí)間	算法	執(zhí)行時(shí)間
ROLD-EPR	5.6	WIN5RB	22.8
ASWM	86.3	LSM-NLR	257.2
ROR-NLM	43.1	RED-Net	5.3
WCSR	1085.1	S-DnCNN	4.1
ALOHA	1875.2	D-DnCNN	5.3

下載: 導(dǎo)出CSV

參考文獻(xiàn)(20)

馬濟(jì)通, 邱天爽, 李蓉, 等. 脈沖噪聲下基于Renyi熵的分?jǐn)?shù)低階雙模盲均衡算法[J]. 電子與信息學(xué)報(bào), 2018, 40(2): 378–385. doi: 10.11999/JEIT170366

MA Jitong, QIU Tianshuang, LI Rong, et al. Dual-mode blind equalization algorithm based on Renyi entropy and fractional lower order statistics under impulsive noise[J]. Journal of Electronics &Information Technology, 2018, 40(2): 378–385. doi: 10.11999/JEIT170366

徐少平, 張貴珍, 李崇禧, 等. 基于深度置信網(wǎng)絡(luò)的隨機(jī)脈沖噪聲快速檢測(cè)算法[J]. 電子與信息學(xué)報(bào), 2019, 41(5): 1130–1136. doi: 10.11999/JEIT180558

XU Shaoping, ZHANG Guizhen, LI Chongxi, et al. A fast random-valued impulse noise detection algorithm based on deep belief network[J]. Journal of Electronics &Information Technology, 2019, 41(5): 1130–1136. doi: 10.11999/JEIT180558

GARNETT R, HUEGERICH T, CHUI C, et al. A universal noise removal algorithm with an impulse detector[J]. IEEE Transactions on Image Processing, 2005, 14(11): 1747–1754. doi: 10.1109/TIP.2005.857261

DONG Yiqiu, CHAN R H, and XU Shufang. A detection statistic for random-valued impulse noise[J]. IEEE Transactions on Image Processing, 2007, 16(4): 1112–1120. doi: 10.1109/TIP.2006.891348

ROY A, SINGHA J, DEVI S S, et al. Impulse noise removal using SVM classification based fuzzy filter from gray scale images[J]. Signal Processing, 2016, 128: 262–273. doi: 10.1016/j.sigpro.2016.04.007

SOLEIMANY S and HAMGHALAM M. A novel random-valued impulse noise detector based on MLP neural network classifier[C]. 2017 Artificial Intelligence and Robotics, Qazvin, Iran, 2017: 165–169. doi: 10.1109/RIOS.2017.7956461.

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. doi: 10.1109/TIP.2017.2662206

ZHANG Kai, ZUO Wangmeng, and 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. doi: 10.1109/TIP.2018.2839891

GUO Shi, YAN Zifei, ZHANG Kai, et al. Toward convolutional blind denoising of real photographs[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 1712–1722. doi: 10.1109/CVPR.2019.00181.

YU Hancheng, ZHAO Li, and WANG Haixian. An efficient procedure for removing random-valued impulse noise in images[J]. IEEE Signal Processing Letters, 2008, 15: 922–925. doi: 10.1109/LSP.2008.2005051

SUN Jiande, LIU Xiaocui, WAN Wenbo, et al. Video hashing based on appearance and attention features fusion via DBN[J]. Neurocomputing, 2016, 213: 84–94. doi: 10.1016/j.neucom.2016.05.098

AKKOUL S, LEDEE R, LECONGE R, et al. A new adaptive switching median filter[J]. IEEE Signal Processing Letters, 2010, 17(6): 587–590. doi: 10.1109/LSP.2010.2048646

XIONG Bo and YIN Zhouping. A universal denoising framework with a new impulse detector and nonlocal means[J]. IEEE Transactions on Image Processing, 2012, 21(4): 1663–1675. doi: 10.1109/TIP.2011.2172804

CHEN C L P, LIU Licheng, CHEN Long, et al. Weighted couple sparse representation with classified regularization for impulse noise removal[J]. IEEE Transactions on Image Processing, 2015, 24(11): 4014–4026. doi: 10.1109/TIP.2015.2456432

JIN K H and YE J C. Sparse and low-rank decomposition of a Hankel structured matrix for impulse noise removal[J]. IEEE Transactions on Image Processing, 2018, 27(3): 1448–1461. doi: 10.1109/TIP.2017.2771471

LIU Peng and FANG Ruogu. Learning pixel-distribution prior with wider convolution for image denoising[EB/OL]. https://arxiv.org/abs/1707.09135, 2017.

MAO Xiaojiao, SHEN Chunhua, and YANG Yubin. Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections[C]. The 30th International Conference on Neural Information Processing Systems, Barcelona, Spain, 2016: 2810–2818.

HUANG Tao, DONG Weisheng, XIE Xuemei, et al. Mixed noise removal via Laplacian scale mixture modeling and nonlocal low-rank approximation[J]. IEEE Transactions on Image Processing, 2017, 26(7): 3171–3186. doi: 10.1109/TIP.2017.2676466

ARBELáEZ P, MAIRE M, FOWLKES C, et al. Contour detection and hierarchical image segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(5): 898–916. doi: 10.1109/TPAMI.2010.161

YUE Zongsheng, YONG Hongwei, ZHAO Qian, et al. Variational denoising network: Toward blind noise modeling and removal[EB/OL]. https://arxiv.org/abs/1908.11314v1, 2019.

相關(guān)文章

施引文獻(xiàn)

資源附件(0)

訪問統(tǒng)計(jì)