CN111127356A - Image blind denoising system - Google Patents

Image blind denoising system Download PDF

Info

Publication number
CN111127356A
CN111127356A CN201911307908.6A CN201911307908A CN111127356A CN 111127356 A CN111127356 A CN 111127356A CN 201911307908 A CN201911307908 A CN 201911307908A CN 111127356 A CN111127356 A CN 111127356A
Authority
CN
China
Prior art keywords
image
blind
network module
denoising
noise
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201911307908.6A
Other languages
Chinese (zh)
Inventor
谢翔
邹少锋
李国林
王志华
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Shenzhen International Graduate School of Tsinghua University
Original Assignee
Tsinghua University
Shenzhen International Graduate School of Tsinghua University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tsinghua University, Shenzhen International Graduate School of Tsinghua University filed Critical Tsinghua University
Priority to CN201911307908.6A priority Critical patent/CN111127356A/en
Publication of CN111127356A publication Critical patent/CN111127356A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30168—Image quality inspection

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Processing (AREA)

Abstract

The invention relates to an image blind denoising system, which comprises: and the blind denoising network module is used for removing the noise of the input image, and the blind denoising network module directly reconstructs the image with the noise by using the pre-training model so as to reduce the iteration times required for generating the optimal reconstructed image. Adopting a coder-decoder structure with jump connection, only using Gaussian white noise as network input, using a noisy image as a reference image, and using a mean square error as a loss function; and the blind image quality evaluation network module is used for evaluating the noise image reconstructed by the blind denoising network module, determining when to terminate the iterative process of the blind denoising network module, and selecting the reconstructed image with the highest score as the final denoising image.

Description

Image blind denoising system
Technical Field
The invention relates to an image blind denoising system, and belongs to the technical field of computer vision processing.
Background
Blind denoising of real noise images is a very important topic, and is more challenging and practical than removal of additive white gaussian noise. The method for realizing blind denoising comprises a method based on traditional image processing and a method based on a convolutional neural network. Noise Client (NC) is a traditional blind denoising algorithm that first estimates a Noise model of an image and then denoises through a multi-scale adaptive non-local bayesian algorithm. Neat Image (NI) is a piece of software issued by ABSoft that eliminates noise in low light photographs taken at high sensitivity settings. Experiments show that the methods still have certain limitation on removing real noisy images. On the other hand, so far, only a small number of researchers have developed blind denoising methods based on Convolutional Neural Networks (CNN). Proposed by ulianov D et al (Deep image prior, ulianov D et al, Proceedings of the IEEE Conference on Computer Vision and pattern recognition.9446-9454, 2018) is a Depth Image Prior (DIP) algorithm which uses a convolutional neural network with randomly initialized parameters, takes gaussian white noise as input, and can implement blind denoising of real images through thousands of iterative operations. However, this method requires thousands of iterative reconstructions and cannot determine the number of iterations required by the network to reconstruct a clean image, and the iterative process must be manually terminated.
Disclosure of Invention
In view of the above-mentioned deficiencies of the prior art, the present invention provides an image blind denoising system, which optimizes a DIP method through transfer learning, determines the number of iterations required by the DIP method through blind image quality evaluation, and reduces the number of iterations required to generate an optimal reconstructed image.
In order to achieve the above object, the present invention provides an image blind denoising system, including: the blind denoising network module is used for removing noise of an input image, comprises a pre-training model, and directly reconstructs a noise image by network weight of a recovered image containing denoised image information obtained after multiple iterations in the pre-training model according to a transfer learning method; and the blind image quality evaluation network module is used for evaluating the noise image reconstructed by the blind denoising network module, determining when to terminate the iterative process of the blind denoising network module, and selecting the reconstructed image with the highest score as the final denoising image.
Further, in the pre-training model, the blind denoising network module generates a reconstructed image in each iteration, and the blind image quality evaluation network module estimates the optimal quality of the reconstructed image by adopting a basic classification structure to determine when to terminate the iteration process of the blind denoising network module.
Furthermore, the blind image quality evaluation network module is connected with a low-pass filter, and the quality score of the reconstructed image estimated by the blind image quality evaluation network module is filtered by the low-pass filter, so that the denoised image with the best quality is obtained.
Further, the blind denoising network module adopts a coder-decoder network structure with jump connection, and the loss function adopts mean square error.
Further, a real noise data set is used for generating a reconstructed noise image through a blind denoising network module, the peak signal-to-noise ratio of the reconstructed noise image is calculated, and the peak signal-to-noise ratio is converted into the distribution of a histogram to be used as the reconstructed image quality score of the blind image quality evaluation network module.
Further, the reconstructed image quality score may be expressed as
Figure BDA0002323667580000021
Here siRepresents the ith fractional interval, N represents the number of fractional intervals,
Figure BDA0002323667580000022
representing the image quality score and the probability of falling into the ith interval, the peak signal-to-noise ratio of each reconstructed image may be linearly converted to μ in order to convert the peak signal-to-noise ratio into a corresponding quality score distribution, the conversion formula being as follows:
Figure BDA0002323667580000023
min (psnr) and max (psnr) respectively represent the minimum and maximum peak signal-to-noise ratios in all reconstructed images, and a normal distribution with μ as a mean, a variance σ of 1.5, and the number of sampling points M is established.
Further, the number of sampling points falling into the ith interval in the normal distribution is calculated, so that the corresponding probability distribution is obtained, and the probability distribution is expressed as
Figure BDA0002323667580000024
When M → ∞ is reached,
Figure BDA0002323667580000025
further, the quality score of the reconstructed image is defined as
Figure BDA0002323667580000026
Further, the blind image quality evaluation network module uses the EMD distance as a loss function, and the specific form is as follows:
Figure BDA0002323667580000027
wherein p and
Figure BDA0002323667580000028
representing true and estimated probability distributions, CDF, respectivelyp(k) Expressed as cumulative probability distribution
Figure BDA0002323667580000029
Due to the adoption of the technical scheme, the invention has the following advantages: aiming at the problem that a real noisy image needs to be denoised under the conditions that a real noisy data set cannot be obtained and a noise model cannot be determined in reality, the blind image denoising system based on the convolutional neural network optimizes a DIP (dual in-line) method through transfer learning so as to reduce the number of iterations required for generating an optimal reconstructed image, and estimates a reconstructed image quality score by designing a blind image quality evaluation network based on the MobileNet so as to obtain an optimal denoising effect. The system has good noise suppression capability on the real image with noise and the denoising effect on the real image data set with noise is superior to the most advanced algorithm in the prior art.
Drawings
FIG. 1 is a schematic structural diagram of a blind image denoising system according to an embodiment of the present invention;
FIG. 2 is a schematic structural diagram of a blind denoising network module according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of a blind image quality evaluation network module according to an embodiment of the present invention.
Formula of concrete implementation
The present invention is described in detail below with reference to the attached drawings. It is to be understood, however, that the drawings are provided solely for the purposes of promoting an understanding of the invention and that they are not to be construed as limiting the invention. In describing the present invention, it is to be understood that the terminology used is for the purpose of description only and is not intended to be indicative or implied of relative importance.
Example one
The embodiment provides an image blind denoising system, as shown in fig. 1, including: and the blind denoising network module is used for removing the noise of the input image, the input of the blind denoising network module is the superposition of Gaussian white noise and uniform noise, and the reference image of the blind denoising network module is a noisy image. In order to reduce the iteration times of image reconstruction, a pre-training model is introduced for a blind denoising network module. According to a transfer learning method, directly reconstructing a noise image by using network weights obtained after multiple rounds of iteration in a pre-training model, wherein the network weights already comprise characteristic information of a de-noised image; and the blind image quality evaluation network module is used for evaluating the noise image reconstructed by the blind denoising network module and determining when to terminate the iterative process of the blind denoising network module. In the pre-training model, a blind denoising network module generates a reconstructed image in each iteration, a blind image quality evaluation network module adopts a basic classification structure to estimate the optimal quality of the reconstructed image so as to determine when to terminate the iteration process of the blind denoising network module, and the reconstructed image with the highest score is selected as the final denoising image. Because the input of the blind denoising network module is Gaussian white noise and uniform noise, the quality of the reconstructed image has obvious local fluctuation, in order to avoid the quality of the reconstructed image falling into a local optimal value and find an optimal quality peak point, the image quality score estimated by the blind image quality evaluation network module is filtered through a low-pass filter, and the signal is smoothed, so that the denoised image with the optimal quality is found. In the embodiment, the system optimizes the DIP method through transfer learning to reduce the number of iterations required to generate the best reconstructed image.
In this embodiment, as shown in fig. 2, the blind denoising network module adopts a coder-decoder network structure with a hopping connection. The input of the blind denoising network module is uniform noise, and Gaussian white noise is added in the iteration process to serve as disturbance. The reference image of the blind denoising network module is a natural image containing obvious noise, and the loss function adopts mean square error (L2). In fig. 2, in the encoder-decoder network structure, five down-sampling blocks, five up-sampling blocks and five jump connection blocks are adopted, the down-sampling blocks are composed of a convolution layer, a down-sampling layer, a batch normalization layer, a leakage activation function, a convolution layer, batch normalization and a leakage activation layer which are sequentially arranged, the up-sampling blocks are composed of a batch normalization layer, a convolution layer, a batch normalization layer, a leakage activation function layer and an up-sampling layer which are sequentially arranged, and the jump connection layers are composed of a convolution layer, a batch normalization layer and a leakage activation function layer which are sequentially arranged. n isu[i]，nd[i]，ns[i]Corresponding to the number of filters for sampling, down-sampling and hopping connections at the i-th layer, respectively. k is a radical ofu[i]，kd[i]，ks[i]Corresponding to the size of the respective convolution kernels, n is, in the present embodiment, the aboveu[i]，nd[i]，ns[i]，ku[i]，kd[i]，ks[i]Preferably: n isu＝nd＝[128,128,128,128,128]，ku＝kd＝[3,3,3,3,3]，ns＝[4,4,4,4,4]，ks＝[1,1,1,1,1]。nu[i]，nd[i]，ns[i]，ku[i]，kd[i]，ks[i]The setting may be set according to specific needs, and is not limited to the above setting. The DIP-based method, which uses a randomly initialized network to fit a noisy image, requires thousands of iterations to obtain an optimal reconstructed image of the noisy image. Inspired by transfer learning, in the embodiment, the noise image is directly reconstructed by adopting the network weight of the recovered image containing the information of the denoised image after multiple iterations of the pre-trained model, so that the iteration times required for achieving the optimal reconstructed image can be greatly reduced.
In this embodiment, as shown in fig. 3, the blind image quality evaluation network module predicts the histogram quality score distribution of the reconstructed image, instead of classifying the image into a high score and a low score or regressing to predict a certain score. The blind image quality evaluation network module adopts MobileNet as a classifier framework, and the MobileNet uses deep separable convolution to construct a lightweight deep neural network. The invention uses MobileNet, and can also replace classification networks such as VGG, inclusion, ResNet and the like. The present invention replaces the original fully-connected layer with one fully-connected layer with 10 neurons and activates with softmax. Since most blind image quality assessment methods use common data sets such as TID, AVA, etc., these data sets have very different attributes from the reconstructed images generated by the present invention. In order to train the blind image quality evaluation network module, a series of reconstructed images are generated by using a noise data set of the real world through the blind denoising network module, and according to the corresponding clean images, the peak signal-to-noise ratio of the reconstructed images is calculated and converted into the distribution of a histogram to be used as the quality score of the reconstructed images.
The quality distribution of the reconstructed image can be expressed as
Figure BDA0002323667580000041
Here siRepresents the ith fractional interval, N represents the number of fractional intervals,
Figure BDA0002323667580000042
representing the sum of the image quality scores and the probability of falling into the ith interval, where s1＜si＜snSetting N as 10, s1＝0.5，s109.5. To convert the peak signal-to-noise ratio into a corresponding mass fraction distribution, the peak signal-to-noise ratio of each reconstructed image may be linearly converted into μ, as follows:
Figure BDA0002323667580000043
here min (psnr) and max (psnr) represent the minimum and maximum peak signal-to-noise ratios in all reconstructed images, respectively. Then, a normal distribution with μ as the mean, the variance σ of 1.5, and the number of sampling points M is established. And finally, calculating the number of sampling points falling into the ith interval, thereby obtaining the corresponding probability distribution. The probability distribution can be expressed as
Figure BDA0002323667580000044
When M → ∞ is reached,
Figure BDA0002323667580000045
after the training of the blind image quality evaluation network module is finished, the network can be used for estimating the histogram distribution of the quality scores of the reconstructed images, and the quality scores representing the reconstructed images are defined as
Figure BDA0002323667580000046
The blind image quality evaluation network module uses an EMD Distance (Earth Mover's Distance) as a loss function, and EMD loss penalizes misclassification according to the Distance of actual classification and prediction classification, which can be specifically expressed as:
Figure BDA0002323667580000051
where p and
Figure BDA0002323667580000052
representing true and estimated probability distributions, CDF, respectivelyp(k) Expressed as cumulative probability distribution
Figure BDA0002323667580000053
Example two
The embodiment is an example of a specific implementation process of the image blind denoising system in the first embodiment, and compares the image blind denoising system with an image blind denoising system in the prior art. In the embodiment, two real noisy image data sets, namely a PolyU data set and a Nam data set, are used, wherein the PolyU data set is used for training a blind image quality evaluation network module, and the Nam data set is used for verifying the denoising performance of the image blind denoising system in the embodiment.
The Nam data set contains noisy images of 11 static scenes, each of which was acquired by taking 500 shots with the same camera and the same camera parameters. The sum and average of 500 images yields a mean image that can be considered as a true clean image. The PolyU data set contains the real noisy images of 40 static scenes, consistent with the method for making the Nam data set, and finally cuts out 100 images with the size of 512 × 512.
The training data of the blind image quality evaluation network module is generated by the DBN module through the PolyU data set. Firstly, selecting images in a PolyU data set, iterating each image for 8000 times by using a DBN module, storing one image every 50 times, and finally obtaining 16000 reconstructed images and a peak signal-to-noise ratio corresponding to the reconstructed images.
The blind image quality assessment network module is trained by first initializing the MobileNet using the pre-trained weights obtained on ImageNet and randomly initializing the weights of its fully connected layers. In the first stage of training, the convolutional layer of MobileNet was frozen and 10 iterations were performed using a learning rate of 0.001. In the second stage of training, the bottom convolutional layer is still frozen and the last convolutional layer is thawed to prevent overfitting. The second stage of training uses a learning rate of 0.001 and 20 iterations are performed. The two stages of training input images are rescaled to 256 x 256. Since the data set is relatively small, the present invention randomly crops the image into 224 x 224 and randomly flips horizontally to avoid overfitting.
In the testing stage, a 12 th image is selected from the Nam data set to pre-train the blind denoising network module, the pre-trained network weight is used for directly reconstructing the rest images of the Nam data set, each image is iterated 5000 times, one reconstructed image is stored every 50 times, the blind image quality evaluation network module is used for carrying out image quality evaluation to obtain scores, the reconstructed image with the highest score is selected as a final denoising image, and the PSNR is calculated according to the corresponding clean image in the data set.
This embodiment and the most advanced denoising system: noise Client (NC), Heat Image (NI), CBM3D and CC were compared. Table 1 shows the comparison of the denoising results of the Image blind denoising system, NoiseClinic (NC), New Image (NI), CBM3D and CC denoising system in this embodiment. The present embodiment obtained the highest PSNR value in 14 out of 15 images of the Nam dataset. This embodiment is superior to the other methods by 0.68dB in terms of average PSNR. In addition, the iteration number required by the reconstructed image to reach the highest PSNR is reduced by 36% by using the blind denoising network module of the pre-training model, and the reduction of the PSNR can be almost ignored, so that the blind denoising network module can save more than one third of time in the process of reconstructing the denoised image.
Table 1 shows the comparison of the denoising results of the CBM3D, NC, NI and CC denoising systems and the image blind denoising system in the embodiment.
Figure BDA0002323667580000061
The above description is only for the specific embodiments of the present application, but the scope of the present application is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present application, and shall be covered by the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims (10)

1. An image blind denoising system, comprising:
the blind denoising network module is used for removing noise of an input image, comprises a pre-training model, and directly reconstructs a noise image by network weight of a recovered image containing denoised image information obtained after multiple iterations in the pre-training model according to a transfer learning method;
and the blind image quality evaluation network module is used for evaluating the noise image reconstructed by the blind denoising network module, determining when to terminate the iterative process of the blind denoising network module, and selecting the reconstructed image with the highest score as the final denoising image.
2. The system of claim 1, wherein in the pre-trained model, the blind denoising network module generates a reconstructed image in each iteration, and the blind image quality estimation network module estimates the best quality of the reconstructed image by using a basic classification structure to decide when to terminate the iteration process of the blind denoising network module.
3. The blind image denoising system of claim 2, wherein the blind image quality evaluation network module is connected to a low pass filter, and the quality score of the reconstructed image estimated by the blind image quality evaluation network module is filtered by the low pass filter, so as to obtain a denoised image with the best quality.
4. The image blind denoising system of any one of claims 1-3, wherein the blind denoising network module employs a coder-decoder network structure with skip connection, the coder-decoder network structure with skip connection is composed of a number of downsample blocks, upsample blocks and skip connection blocks.
5. The system as claimed in any one of claims 1 to 3, wherein the blind denoising network module input is uniform noise, and Gaussian white noise is added as disturbance in the iterative process, the reference image of the blind denoising network module is a natural image containing significant noise, and the loss function of the blind denoising network module is mean square error.
6. The image blind denoising system of any one of claims 1-3, wherein a reconstructed noise image is generated by the blind denoising network module using a true noise data set, and a peak signal-to-noise ratio of the reconstructed noise image is calculated, and the peak signal-to-noise ratio is converted into a distribution of a histogram as a reconstructed image quality score of the blind image quality assessment network module.
7. The system for blind denoising of images of claim 6, wherein the reconstructed image quality score is representable as
Figure FDA0002323667570000011
Here siRepresents the ith fractional interval, N represents the number of fractional intervals,
Figure FDA0002323667570000012
representing the image quality score and the probability of falling into the ith interval, the peak signal-to-noise ratio of each reconstructed image may be linearly converted to μ in order to convert the peak signal-to-noise ratio into a corresponding quality score distribution, the conversion formula being as follows:
Figure FDA0002323667570000021
min (psnr) and max (psnr) respectively represent the minimum and maximum peak signal-to-noise ratios in all reconstructed images, and a normal distribution with μ as a mean, a variance σ of 1.5, and the number of sampling points M is established.
8. The blind image denoising system of claim 7, wherein the number of samples falling in the ith interval in the normal distribution is calculated, thereby obtaining a corresponding probability distribution, and the probability distribution tableShown as
Figure FDA0002323667570000022
Figure FDA0002323667570000023
When M → ∞ is reached,
Figure FDA0002323667570000024
9. the system for blind denoising of images of claim 8, wherein the quality score of the reconstructed image is defined as
Figure FDA0002323667570000025
10. The system for blind denoising of images according to any one of claims 7-9, wherein the blind image quality assessment network module uses EMD distance as a loss function in the form of:
Figure FDA0002323667570000026
wherein p and
Figure FDA0002323667570000027
representing true and estimated probability distributions, CDF, respectivelyp(k) Expressed as cumulative probability distribution
Figure FDA0002323667570000028
CN201911307908.6A 2019-12-18 2019-12-18 Image blind denoising system Pending CN111127356A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911307908.6A CN111127356A (en) 2019-12-18 2019-12-18 Image blind denoising system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911307908.6A CN111127356A (en) 2019-12-18 2019-12-18 Image blind denoising system

Publications (1)

Publication Number Publication Date
CN111127356A true CN111127356A (en) 2020-05-08

Family

ID=70499511

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911307908.6A Pending CN111127356A (en) 2019-12-18 2019-12-18 Image blind denoising system

Country Status (1)

Country Link
CN (1) CN111127356A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112634159A (en) * 2020-12-23 2021-04-09 中国海洋大学 Hyperspectral image denoising method based on blind noise estimation
CN112927159A (en) * 2021-03-11 2021-06-08 清华大学深圳国际研究生院 True image denoising method based on multi-scale selection feedback network
CN113476064A (en) * 2021-07-23 2021-10-08 浙江大学 Single-scanning double-tracer PET signal separation method based on BCD-ED
CN114663307A (en) * 2022-03-22 2022-06-24 哈尔滨工业大学 Integrated image denoising system based on uncertainty network
CN115482162A (en) * 2022-09-02 2022-12-16 华南理工大学 Implicit image blind denoising method based on random rearrangement and label-free model

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180240219A1 (en) * 2017-02-22 2018-08-23 Siemens Healthcare Gmbh Denoising medical images by learning sparse image representations with a deep unfolding approach
CN109035163A (en) * 2018-07-09 2018-12-18 南京信息工程大学 A kind of adaptive denoising method based on deep learning
CN109727663A (en) * 2018-09-30 2019-05-07 上海联影医疗科技有限公司 The system and method for generating the neural network model for image procossing
CN110197183A (en) * 2019-04-17 2019-09-03 深圳大学 A kind of method, apparatus and computer equipment of Image Blind denoising

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180240219A1 (en) * 2017-02-22 2018-08-23 Siemens Healthcare Gmbh Denoising medical images by learning sparse image representations with a deep unfolding approach
CN109035163A (en) * 2018-07-09 2018-12-18 南京信息工程大学 A kind of adaptive denoising method based on deep learning
CN109727663A (en) * 2018-09-30 2019-05-07 上海联影医疗科技有限公司 The system and method for generating the neural network model for image procossing
CN110197183A (en) * 2019-04-17 2019-09-03 深圳大学 A kind of method, apparatus and computer equipment of Image Blind denoising

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SHAOFENG ZOU等: "A CNN-Based Blind Denoising Method for Endoscopic Images", 《IEEE》, pages 2 - 4 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112634159A (en) * 2020-12-23 2021-04-09 中国海洋大学 Hyperspectral image denoising method based on blind noise estimation
CN112634159B (en) * 2020-12-23 2022-07-26 中国海洋大学 Hyperspectral image denoising method based on blind noise estimation
CN112927159A (en) * 2021-03-11 2021-06-08 清华大学深圳国际研究生院 True image denoising method based on multi-scale selection feedback network
CN112927159B (en) * 2021-03-11 2022-08-02 清华大学深圳国际研究生院 True image denoising method based on multi-scale selection feedback network
CN113476064A (en) * 2021-07-23 2021-10-08 浙江大学 Single-scanning double-tracer PET signal separation method based on BCD-ED
CN113476064B (en) * 2021-07-23 2023-09-01 浙江大学 BCD-ED-based single-scanning double-tracer PET signal separation method
CN114663307A (en) * 2022-03-22 2022-06-24 哈尔滨工业大学 Integrated image denoising system based on uncertainty network
CN115482162A (en) * 2022-09-02 2022-12-16 华南理工大学 Implicit image blind denoising method based on random rearrangement and label-free model
CN115482162B (en) * 2022-09-02 2023-10-24 华南理工大学 Implicit image blind denoising method based on random rearrangement and label-free model

Similar Documents

Publication Publication Date Title
CN111127356A (en) Image blind denoising system
CN109360156B (en) Single image rain removing method based on image block generation countermeasure network
CN108765319B (en) Image denoising method based on generation countermeasure network
CN108876735B (en) Real image blind denoising method based on depth residual error network
CN106204467B (en) Image denoising method based on cascade residual error neural network
CN111028163A (en) Convolution neural network-based combined image denoising and weak light enhancement method
CN105844627B (en) A kind of sea-surface target image background suppressing method based on convolutional neural networks
Chou et al. Turbulent-PSO-based fuzzy image filter with no-reference measures for high-density impulse noise
CN111861894A (en) Image motion blur removing method based on generating type countermeasure network
CN110349112B (en) Two-stage image denoising method based on self-adaptive singular value threshold
CN113592715B (en) Super-resolution image reconstruction method for small sample image set
CN112488934B (en) CS-TCGAN-based finger vein image denoising method
Dementyiev et al. Use of images augmentation and implementation of doubly stochastic models for improving accuracy of recognition algorithms based on convolutional neural networks
Paul et al. Modified convolutional neural network with pseudo-CNN for removing nonlinear noise in digital images
CN111754447A (en) Infrared and visible light image fusion method based on multi-state context hidden Markov model
CN112950505B (en) Image processing method, system and medium based on generation countermeasure network
Wang et al. Multi-feature fusion tracking algorithm based on generative compression network
CN114331913B (en) Motion blurred image restoration method based on residual attention block
CN113744148B (en) Method and system for building and denoising inscription handwriting image denoising model
CN115358952A (en) Image enhancement method, system, equipment and storage medium based on meta-learning
ZhiPing et al. A new generative adversarial network for texture preserving image denoising
CN111047537A (en) System for recovering details in image denoising
CN114862699B (en) Face repairing method, device and storage medium based on generation countermeasure network
CN113034475B (en) Finger OCT (optical coherence tomography) volume data denoising method based on lightweight three-dimensional convolutional neural network
CN115330639A (en) Depth-enhanced image denoising method based on non-local attention

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20200508

WD01 Invention patent application deemed withdrawn after publication