CN112991199A - Image high-low frequency decomposition noise removing method based on residual error dense network - Google Patents
Image high-low frequency decomposition noise removing method based on residual error dense network Download PDFInfo
- Publication number
- CN112991199A CN112991199A CN202110182091.5A CN202110182091A CN112991199A CN 112991199 A CN112991199 A CN 112991199A CN 202110182091 A CN202110182091 A CN 202110182091A CN 112991199 A CN112991199 A CN 112991199A
- Authority
- CN
- China
- Prior art keywords
- image
- noise
- frequency
- low
- img
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 59
- 238000000354 decomposition reaction Methods 0.000 title claims abstract description 20
- 238000012549 training Methods 0.000 claims abstract description 72
- 238000000605 extraction Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 description 4
- 238000003384 imaging method Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000003709 image segmentation Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000001815 facial effect Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
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
- G06N3/08—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
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
Abstract
The invention discloses an image high-low frequency decomposition noise removing method based on a residual error dense network, which specifically comprises the following steps: step 1, acquiring an image to be trained; step 2, Gaussian noise is added to the image to be trained in the step 1, and an image pair X is constructed; step 3, respectively using the image to be trained in the step 1 and the noise data set obtained in the step 2 to obtain corresponding high-frequency image training samples and low-frequency image training samples by using a high-pass filter; step 4, respectively inputting the high-frequency image training sample and the low-frequency image training sample obtained in the step 3 into a residual dense network for training, and respectively obtaining a high-frequency image with noise removed and a low-frequency image with noise removed; and 5, adding the high-frequency image and the low-frequency image which are obtained in the step 4 and subjected to noise removal in a one-to-one correspondence mode to obtain a denoised whole image. The method solves the problems that edge information is fuzzy and artifacts are easy to generate after image denoising in the prior art.
Description
Technical Field
The invention belongs to the technical field of image processing methods, and relates to an image high-low frequency decomposition noise removing method based on a residual error dense network.
Background
Due to the influence of various factors such as imaging equipment, an image is interfered by noise in the imaging or sensing process, so that subsequent tasks such as image segmentation and target identification are influenced and cannot be smoothly carried out. For example, when public monitoring equipment is used for determining a criminal suspect, the noise of an image makes it very difficult to distinguish the facial features of the criminal suspect; the imaging pixels of small target objects in the remote sensing image are fewer, and the small target objects in the image are not well recognized due to the existence of noise. In view of the above circumstances, how to accurately remove image noise and protect the original image details from being damaged while removing the image noise becomes a problem to be solved urgently.
Disclosure of Invention
The invention aims to provide an image high-low frequency decomposition noise removing method based on a residual error dense network, and solves the problems that in the prior art, edge information is fuzzy after image denoising and artifacts are prone to being generated.
The invention adopts the technical scheme that the image high-low frequency decomposition noise removing method based on the residual error dense network specifically comprises the following steps:
step 1, acquiring an image to be trained;
step 4, respectively inputting the high-frequency image training sample and the low-frequency image training sample obtained in the step 3 into a residual dense network for training, and respectively obtaining a high-frequency image with noise removed and a low-frequency image with noise removed;
and 5, adding the high-frequency image subjected to noise removal and the low-frequency image subjected to noise removal obtained in the step 4 in a one-to-one correspondence manner to obtain a denoised whole image.
The invention is also characterized in that:
the specific process of the step 1 is as follows:
selecting A training sample images to form a training sample image set, randomly selecting m images from the training sample image set, and randomly cutting i pictures with the size of n multiplied by n from each picture to obtain m multiplied by i pictures with the size of n multiplied by n.
The specific process of the step 2 is as follows:
step 2.1, copying the m multiplied by i pictures processed in the step 1, and adding the same Gaussian noise to each copied picture to obtain an artificial noise picture;
and 2.2, the images without noise in the step 1 and the images with noise in the step 2.1 are in one-to-one correspondence to form m × i image pairs, wherein the m × i image pairs X are { noise _ img, clean _ img }, and the noise _ img and the clean _ img respectively represent the images with noise and the images without noise.
The specific process of adding gaussian noise in step 2.1 is as follows:
step A, the probability density of Gaussian noise follows Gaussian distribution, as shown in formula (1):
wherein u is mean and sigma is standard deviation sigma;
step B, setting parameters mean to be 0 and sigma to be 30, and generating a Gaussian random number according to Gaussian distribution, namely a formula (1);
step C, adding the random number which is in accordance with Gaussian distribution in the step B to each pixel in each copied image to obtain an output pixel, namely the output pixel is the input pixel plus the Gaussian random number, and limiting or scaling the value of the output pixel between [ 0-255 ];
and D, repeating the steps B to C until all pixels of the whole image without the noise are circulated, and obtaining an output image, wherein the output image is an artificial noise image.
The specific process of the step 3 is as follows:
step 3.1, obtaining m × i high-frequency image pairs X by passing m × i image pairs X ═ { noised _ img, clean _ img } in step 2.2 through a high-pass filterhighHigh-frequency image training samples, where non _ img _ high and clean _ img _ high represent the high-frequency image of the noise image and the high-frequency image without noise, respectively；
Step 3.2, matching the m × i image pairs X ═ n _ img, clear _ img in step 2.2 with the m × i high-frequency image pairs X obtained in step 3.1highSubtracting { noised _ img _ high, clean _ img _ high } in one-to-one correspondence to obtain m × i low-frequency image pairs XlowAnd { noise _ img _ low, clean _ img _ low }, i.e., low-frequency image training samples, wherein noise _ img _ low and clean _ img _ low represent low-frequency images of a noisy image and a non-noisy image, respectively.
The specific process of the step 4 is as follows:
step 4.1, inputting the high-frequency image training sample into a residual error dense network for training to obtain a high-frequency image with noise removed;
and 4.2, inputting the low-frequency image training sample into a residual error dense network for training to obtain a low-frequency image with noise removed.
The specific process of the step 4.1 is as follows:
step 4.1.1, m × i high-frequency images in step 3.1 are paired with XhighInputting the obtained signals into a residual error dense network for feature extraction to obtain spatial features of the high-frequency training image;
step 4.1.2, inputting the spatial characteristics of the high-frequency image obtained in the step 4.1.1 into a dense block of a residual dense network for training, and setting an image loss function as a pixel-by-pixel loss function MSElossWhen the loss function reaches a minimum, the high-frequency image is output as shown in the following equation (2):
wherein x represents a high-frequency noise image, y represents a high-frequency image without noise, F (x) represents a denoised high-frequency image obtained by the high-frequency noise image through residual error dense network training, and C, W, H represents the channel, width and height of the high-frequency image pair (x, y) respectively;
and 4.1.3, reconstructing the high-frequency image output in the step 4.1.2 from the space characteristic through a deconvolution layer of a residual dense network to the image characteristic, and obtaining the denoised high-frequency image.
The specific process of the step 4.2 is as follows:
step 4.2.1, m × i low-frequency images in step 3.2 are paired with XlowInputting the obtained result into a residual error dense network for feature extraction to obtain spatial features of the low-frequency training image;
step 4.2.2, inputting the spatial characteristics of the low-frequency image obtained in the step 4.2.1 into a dense block of a residual dense network for training, and setting an image loss function as a pixel-by-pixel loss function MSEloss /When the loss function reaches a minimum, the low-frequency image is output as shown in the following equation (3):
wherein x ' represents a low-frequency noise image, y ' represents a low-frequency image without noise, F ' (x ') represents a denoised low-frequency image obtained by the low-frequency noise image through residual dense network training, and C ', W ' and H ' represent the channel, width and height of the low-frequency image pair (x ', y ') respectively;
and 4.2.3, reconstructing the low-frequency image output in the step 4.2.2 from the spatial features through a deconvolution layer of a residual dense network back to image features to obtain a denoised low-frequency image.
The invention has the following beneficial effects:
(1) the image noise removing method is mainly based on the residual error dense network for training and image generation, and can remove image noise to the maximum extent and protect the complex edge information of the image;
(2) the image noise removing method can provide a good preprocessing operation for the follow-up research of image segmentation, target detection and identification;
(3) the image noise removing method adopts high-low frequency decomposition, and trains in different degrees aiming at the multiple noise of the high-frequency image and the less noise of the low-frequency image, so that the noise can be removed more thoroughly, and the result is more ideal.
Drawings
FIG. 1 is a general flow chart of the image high and low frequency decomposition noise removing method based on the residual error dense network of the present invention;
FIG. 2 is a structure of a residual dense network in the method for removing high and low frequency decomposition noise of an image based on the residual dense network of the present invention;
FIG. 3 is a network structure diagram of a residual error dense module (RDB) in the image high and low frequency decomposition noise removal method network based on the residual error dense network of the present invention;
FIG. 4 is a denoising map of an embodiment of the image high and low frequency decomposition noise removing method based on the residual error dense network.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses an image high-low frequency decomposition noise removing method based on a residual error dense network, which specifically comprises the following steps as shown in figure 1:
step 1, acquiring an image to be trained;
the specific process of the step 1 is as follows:
selecting A training sample images to form a training sample image set, randomly selecting m images from the training sample image set, and randomly cutting i pictures with the size of n multiplied by n from each picture to obtain m multiplied by i pictures with the size of n multiplied by n.
the specific process of the step 2 is as follows:
step 2.1, copying the m multiplied by i pictures processed in the step 1, and adding the same Gaussian noise to each copied picture to obtain an artificial noise picture;
the specific process of adding gaussian noise in step 2.1 is as follows:
step A, the probability density of Gaussian noise follows Gaussian distribution (normal distribution), and is shown in formula (1):
wherein u is mean and sigma is standard deviation sigma; when noise is added to an image, an operation is performed on all pixels of the image. The image without noise added in the input program is added with a random number conforming to a gaussian distribution for each pixel of the input image, and an output pixel is obtained, that is, the output pixel is equal to the input pixel + the gaussian random number.
Step B, setting parameters mean to be 0 and sigma to be 30, and generating a Gaussian random number according to Gaussian distribution (normal distribution), namely a formula (1);
step C, adding each pixel in each copied image to the random number which accords with Gaussian distribution in the step B to obtain an output pixel, and limiting or scaling the value of the output pixel between 0 and 255;
and D, repeating the steps B to C until all pixels of the whole image without the noise are circulated, and obtaining an output image, wherein the output image is an artificial noise image.
And 2.2, the images without noise in the step 1 and the images with noise in the step 2.1 are in one-to-one correspondence to form m × i image pairs, wherein the m × i image pairs X are { noise _ img, clean _ img }, and the noise _ img and the clean _ img respectively represent the images with noise and the images without noise.
the specific process of the step 3 is as follows:
step 3.1, obtaining m × i high-frequency image pairs X by passing m × i image pairs X ═ { noised _ img, clean _ img } in step 2.2 through a high-pass filterhighThe method comprises the steps of (non _ img _ high, clean _ img _ high), namely high-frequency image training samples, wherein the non _ img _ high and the clean _ img _ high respectively represent a high-frequency image of a noise image and a high-frequency image without noise;
step 3.2, dividing the m × i image pairs X in step 2.2 into X{ noised _ img, clean _ img } pairs of m × i high-frequency images X obtained in step 3.1highSubtracting { noised _ img _ high, clean _ img _ high } in one-to-one correspondence to obtain m × i low-frequency image pairs XlowAnd { noise _ img _ low, clean _ img _ low }, i.e., low-frequency image training samples, wherein noise _ img _ low and clean _ img _ low represent low-frequency images of a noisy image and a non-noisy image, respectively.
Step 4, respectively inputting the high-frequency image training sample and the low-frequency image training sample obtained in the step 3 into a residual dense network for training, and respectively obtaining a high-frequency image with noise removed and a low-frequency image with noise removed;
step 4.1, inputting the high-frequency image training sample into a residual error dense network for training to obtain a high-frequency image with noise removed;
and (3) inputting the noisy high-frequency layer image noised _ img _ high and the non-noisy high-frequency layer image clean _ img _ high into the residual error dense network in the step (3.1). The high-frequency image with noise is calculated by a network, and specifically comprises the following steps: the method comprises the steps of firstly passing through two convolution modules, then passing through 3 residual error dense (RDB) modules, and finally passing through a deconvolution module and a Tanh module, and outputting a denoised clean image with the same size as an input image. The residual dense network is shown in fig. 2, and the specific implementation process is as follows.
And 4.1.1, performing feature extraction on the high-frequency image with the noise through two convolution modules, thereby obtaining the spatial features of the high-frequency training image. The two convolutional layer structures are shown in the first two convolutions in fig. 2, where Conv, k7, n32, s1 denote convolution operations, where the convolution kernel size is 7 × 7, the number of convolution kernels is 32, and the step size s is set to 1; conv, k5, n64 and s2 represent convolution operations, the size of a convolution kernel is 5 multiplied by 5, the number of the convolution kernels is 64, and the step size s is set to be 2; BN represents the Batch Normalizanti normalization operation.
And 4.1.2, inputting the high-frequency image space characteristics obtained in the step 4.1.1 into 3 dense blocks in a residual dense network for training, and reducing the alternation of loss functions to achieve the image denoising effect. The specific structure of the dense block is shown in fig. 3, where ConvBlock1-ConvBlock6 respectively represent 6 convolution operations, Conv, k3, n32, and s1 represent convolution operations, the size of a convolution kernel is 3 × 3, the number of convolution kernels is 32, and the step size s is set to 1. The image feature transfer process in the network structure is represented as formula (2):
Fn＝ReLU(Wn[F0，F1，......，Fn-1])，n∈{1，2，3，4，5，6} (2)；
in the formula (2), F0F1，...Fn-1Each representing the output, W, of each ConvBlocknRepresents the weight parameter in each ConvBlock, ReLU represents a non-linear activation function, and Concat represents F0，F1，...Fn-1Connected with each other in the third channel direction, and subjected to Conv, k1, s1 convolution operation to enable F7And F0The number of channels of (2) is the same, and finally F7And F0The two tensors are added to output F8I.e. the output of one RDB. The loss function used in the invention is a pixel-by-pixel loss function MSElossSpecifically, the calculation is as in formula (3), the size of the loss function can be monitored in the training process, and the loss function gradually decreases with the increase of the training times, but overfitting of the image is caused by excessive training, so that the loss function becomes larger. When the loss function reaches the minimum, the network model required by the invention is obtained, and at the moment, the high-frequency image is output;
wherein x represents a high-frequency noise image, y represents a high-frequency image without noise, f (x) represents a denoised high-frequency image obtained by the high-frequency noise image through residual error dense network training, and C, W, H represents the channel, width and height of the high-frequency image pair (x, y), respectively.
And 4.1.3, reconstructing the space characteristics of the high-frequency image denoised in the step 4.1.2 back to the original image characteristics through deconvolution to obtain the denoised high-frequency image. Specifically, the convolution is implemented by the last two convolution layers in fig. 2, where Conv, k5, n32, s1/2 represent deconvolution operation, the size of deconvolution kernel is 5 × 5, the number of deconvolution kernels is 32, the step size is 1/2, Conv, k7, n3, s1 represent convolution operation, the size of convolution kernel is 7 × 7, the number of convolution kernels is 3, and the step size s is 1; tanh represents a nonlinear activation function.
Step 4.2, inputting the low-frequency image training sample into a residual dense network for training to obtain a low-frequency image with noise removed;
and (3) inputting the noisy low-frequency layer image noised _ img _ low and the noiseless low-frequency layer image clean _ img _ low in the step (3.2) into the residual error dense network. The high-frequency image with noise is calculated by a network, and specifically comprises the following steps: the method comprises the steps of firstly passing through two convolution modules, then passing through 3 residual error dense (RDB) modules, and finally passing through a deconvolution module and a Tanh module, and outputting a denoised clean image with the same size as an input image. The residual dense network is shown in fig. 2, and the specific implementation process is as follows.
And 4.2.1, performing feature extraction on the low-frequency image with the noise through two convolution modules, thereby obtaining the spatial features of the low-frequency training image. The two convolutional layer structures are shown in the first two convolutions in fig. 2, where Conv, k7, n32, s1 denote convolution operations, where the convolution kernel size is 7 × 7, the number of convolution kernels is 32, and the step size s is set to 1; conv, k5, n64 and s2 represent convolution operations, the size of a convolution kernel is 5 multiplied by 5, the number of the convolution kernels is 64, and the step size s is set to be 2; BN represents the Batch Normalizanti normalization operation.
And 4.2.2, inputting the spatial characteristics of the low-frequency image obtained in the step 4.2.1 into 3 dense blocks in the residual dense network for training, and reducing the alternation of the loss function to achieve the image denoising effect. The specific structure of the dense block is shown in fig. 3, where ConvBlock1-ConvBlock6 respectively represent 6 convolution operations, Conv, k3, n32, and s1 represent convolution operations, the size of a convolution kernel is 3 × 3, the number of convolution kernels is 32, and the step size s is set to 1. The image feature transfer process in the network structure is represented as formula (2):
Fn＝ReLU(Wn[F0，F1，......，Fn-1])，n∈{1，2，3，44，5，6} (2)；
in the formula, F0，F1，...Fn-1Each representing the output, W, of each ConvBlocknRepresents the weight parameter in each ConvBlock, ReLU represents a non-linear activation function, and Concat represents F0，F1，...Fn-1Connected with each other in the third channel direction, and subjected to Conv, k1, s1 convolution operation to enable F7And F0The number of channels of (2) is the same, and finally F7And F0The two tensors are added to output F8I.e. the output of one RDB. The loss function we use is the pixel-wise loss function MSEloss /Specifically, the calculation is as in formula (4), the size of the loss function can be monitored in the training process, and the loss function gradually decreases with the increase of the training times, but overfitting of the image is caused by excessive training, so that the loss function becomes larger. When the loss function reaches the minimum, the network model required by the invention is obtained, and the low-frequency image is output:
wherein x ' represents a low-frequency noise image, y ' represents a low-frequency image without noise, F ' (x ') represents a denoised low-frequency image obtained by the low-frequency noise image through residual dense network training, and C ', W ' and H ' represent the channel, width and height of the low-frequency image pair (x ', y ') respectively;
and 4.2.3, reconstructing the space characteristics of the low-frequency image denoised in the step 4.2.2 back to the original image characteristics through deconvolution to obtain the denoised high-frequency image. Specifically, it is implemented by the last two convolutional layers of fig. 2, where Conv, k5, n32, s1/2, represents the deconvolution operation, the size of the deconvolution kernel is 5 × 5, the number of deconvolution kernels is 32, the step size is 1/2, Conv, k7, n3, s1 represents the convolution operation, the size of the convolution kernel is 7 × 7, the number of convolution kernels is 3, and the step size s is 1; tanh represents a nonlinear activation function.
The specific operation in the step 5 is as follows:
and (4) correspondingly adding the denoised high-frequency image and the denoised low-frequency image obtained in the steps 4.1.3 and 4.2.3 one by one to obtain a final denoised image.
And finally, after training data is finished, the loss function reaches the minimum to obtain a trained network, respectively inputting the high-frequency layer test data into the model generated in the step 4.1 to obtain a denoised high-frequency layer image and inputting the low-frequency layer test data into the model generated in the step 4.2 to obtain a denoised low-frequency image according to the test set data preprocessed in the steps 1, 2 and 3. And 5, obtaining the whole de-noised image. Fig. 4 shows a denoised image obtained by processing a noise image through a network.
Claims (8)
1. A method for removing high and low frequency decomposition noise of an image based on a residual error dense network is characterized by comprising the following steps: the method specifically comprises the following steps:
step 1, acquiring an image to be trained;
step 2, Gaussian noise is added to the image to be trained in the step 1, and an image pair X is constructed;
step 3, respectively using the image to be trained in the step 1 and the noise data set obtained in the step 2 to obtain corresponding high-frequency image training samples and low-frequency image training samples by using a high-pass filter;
step 4, respectively inputting the high-frequency image training sample and the low-frequency image training sample obtained in the step 3 into a residual dense network for training, and respectively obtaining a high-frequency image with noise removed and a low-frequency image with noise removed;
and 5, adding the high-frequency image subjected to noise removal and the low-frequency image subjected to noise removal obtained in the step 4 in a one-to-one correspondence manner to obtain a denoised whole image.
2. The method for removing the high and low frequency decomposition noise of the image based on the residual error dense network as claimed in claim 1, wherein: the specific process of the step 1 is as follows:
selecting A training sample images to form a training sample image set, randomly selecting m images from the training sample image set, and randomly cutting i pictures with the size of n multiplied by n from each picture to obtain m multiplied by i pictures with the size of n multiplied by n.
3. The method according to claim 2, wherein the method for removing the noise in the image decomposition comprises: the specific process of the step 2 is as follows:
step 2.1, copying the m multiplied by i pictures processed in the step 1, and adding the same Gaussian noise to each copied picture to obtain an artificial noise picture;
and 2.2, the images without noise in the step 1 and the images with noise in the step 2.1 are in one-to-one correspondence to form m × i image pairs, wherein the m × i image pairs X are { noise _ img, clean _ img }, and the noise _ img and the clean _ img respectively represent the images with noise and the images without noise.
4. The method according to claim 3, wherein the method for removing the noise in the image decomposition comprises: the specific process of adding gaussian noise in step 2.1 is as follows:
step A, the probability density of Gaussian noise follows Gaussian distribution, as shown in formula (1):
wherein u is mean and sigma is standard deviation sigma;
step B, setting parameters mean to be 0 and sigma to be 30, and generating a Gaussian random number according to Gaussian distribution, namely a formula (1);
step C, adding the random number which is in accordance with Gaussian distribution in the step B to each pixel in each copied image to obtain an output pixel, namely the output pixel is the input pixel plus the Gaussian random number, and limiting or scaling the value of the output pixel between [ 0-255 ];
and D, repeating the steps B to C until all pixels of the whole image without the noise are circulated, and obtaining an output image, wherein the output image is an artificial noise image.
5. The method according to claim 3, wherein the method for removing the noise in the image decomposition comprises: the specific process of the step 3 is as follows:
step 3.1, obtaining m × i high-frequency image pairs X by passing m × i image pairs X ═ { noised _ img, clean _ img } in step 2.2 through a high-pass filterhighThe method comprises the steps of (non _ img _ high, clean _ img _ high), namely high-frequency image training samples, wherein the non _ img _ high and the clean _ img _ high respectively represent a high-frequency image of a noise image and a high-frequency image without noise;
step 3.2, matching the m × i image pairs X ═ n _ img, clear _ img in step 2.2 with the m × i high-frequency image pairs X obtained in step 3.1highSubtracting { noised _ img _ high, clean _ img _ high } in one-to-one correspondence to obtain m × i low-frequency image pairs XlowAnd { noise _ img _ low, clean _ img _ low }, i.e., low-frequency image training samples, wherein noise _ img _ low and clean _ img _ low represent low-frequency images of a noisy image and a non-noisy image, respectively.
6. The method according to claim 5, wherein the method for removing the noise in the image decomposition comprises: the specific process of the step 4 is as follows:
step 4.1, inputting the high-frequency image training sample into a residual error dense network for training to obtain a high-frequency image with noise removed;
and 4.2, inputting the low-frequency image training sample into a residual error dense network for training to obtain a low-frequency image with noise removed.
7. The method according to claim 6, wherein the method for removing the noise in the image decomposition comprises: the specific process of the step 4.1 is as follows:
step 4.1.1, m × i high-frequency images in step 3.1 are paired with XhighInputting the parameters { noised _ img _ high, clean _ img _ high } into a residual error dense network for feature extraction, thereby obtaining high-frequency trainingSpatial features of the image;
step 4.1.2, inputting the spatial characteristics of the high-frequency image obtained in the step 4.1.1 into a dense block of a residual dense network for training, and setting an image loss function as a pixel-by-pixel loss function MSElossWhen the loss function reaches a minimum, the high-frequency image is output as shown in the following equation (2):
wherein x represents a high-frequency noise image, y represents a high-frequency image without noise, F (x) represents a denoised high-frequency image obtained by the high-frequency noise image through residual error dense network training, and C, W, H represents the channel, width and height of the high-frequency image pair (x, y) respectively;
and 4.1.3, reconstructing the high-frequency image output in the step 4.1.2 from the space characteristic through a deconvolution layer of a residual dense network to the image characteristic, and obtaining the denoised high-frequency image.
8. The method according to claim 6, wherein the method for removing the noise in the image decomposition comprises: the specific process of the step 4.2 is as follows:
step 4.2.1, m × i low-frequency images in step 3.2 are paired with XlowInputting the obtained result into a residual error dense network for feature extraction to obtain spatial features of the low-frequency training image;
step 4.2.2, inputting the spatial characteristics of the low-frequency image obtained in the step 4.2.1 into a dense block of a residual dense network for training, and setting an image loss function as a pixel-by-pixel loss function MSEloss /When the loss function reaches a minimum, the low-frequency image is output as shown in the following equation (3):
wherein x ' represents a low-frequency noise image, y ' represents a low-frequency image without noise, F ' (x ') represents a denoised low-frequency image obtained by the low-frequency noise image through residual dense network training, and C ', W ' and H ' represent the channel, width and height of the low-frequency image pair (x ', y ') respectively;
and 4.2.3, reconstructing the low-frequency image output in the step 4.2.2 from the spatial features through a deconvolution layer of a residual dense network back to image features to obtain a denoised low-frequency image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110182091.5A CN112991199B (en) | 2021-02-08 | 2021-02-08 | Image high-low frequency decomposition noise removal method based on residual dense network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110182091.5A CN112991199B (en) | 2021-02-08 | 2021-02-08 | Image high-low frequency decomposition noise removal method based on residual dense network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112991199A true CN112991199A (en) | 2021-06-18 |
| CN112991199B CN112991199B (en) | 2024-02-09 |
Family
ID=76393013
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110182091.5A Active CN112991199B (en) | 2021-02-08 | 2021-02-08 | Image high-low frequency decomposition noise removal method based on residual dense network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112991199B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113379638A (en) * | 2021-06-23 | 2021-09-10 | 西安理工大学 | SAR image despeckle method based on wavelet transform and interval dense network |
| CN114129171A (en) * | 2021-12-01 | 2022-03-04 | 山东省人工智能研究院 | Electrocardiosignal noise reduction method based on improved residual error dense network |
| CN114638829A (en) * | 2022-05-18 | 2022-06-17 | 安徽数智建造研究院有限公司 | Anti-interference training method of tunnel lining detection model and tunnel lining detection method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090324117A1 (en) * | 2008-06-27 | 2009-12-31 | Microsoft Corporation | Image denoising techniques |
| CN108805833A (en) * | 2018-05-29 | 2018-11-13 | 西安理工大学 | Miscellaneous minimizing technology of copybook binaryzation ambient noise of network is fought based on condition |
| US20200034948A1 (en) * | 2018-07-27 | 2020-01-30 | Washington University | Ml-based methods for pseudo-ct and hr mr image estimation |
| CN110838089A (en) * | 2019-09-18 | 2020-02-25 | 浙江工业大学 | Fast image denoising method based on OctBlock dense block |
| CN112270646A (en) * | 2020-11-05 | 2021-01-26 | 浙江传媒学院 | Super-resolution enhancement method based on residual error dense jump network |
-
2021
- 2021-02-08 CN CN202110182091.5A patent/CN112991199B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090324117A1 (en) * | 2008-06-27 | 2009-12-31 | Microsoft Corporation | Image denoising techniques |
| CN108805833A (en) * | 2018-05-29 | 2018-11-13 | 西安理工大学 | Miscellaneous minimizing technology of copybook binaryzation ambient noise of network is fought based on condition |
| US20200034948A1 (en) * | 2018-07-27 | 2020-01-30 | Washington University | Ml-based methods for pseudo-ct and hr mr image estimation |
| CN110838089A (en) * | 2019-09-18 | 2020-02-25 | 浙江工业大学 | Fast image denoising method based on OctBlock dense block |
| CN112270646A (en) * | 2020-11-05 | 2021-01-26 | 浙江传媒学院 | Super-resolution enhancement method based on residual error dense jump network |
Non-Patent Citations (3)
| Title |
|---|
| 吴从中;陈曦;季栋;詹曙;: "结合深度残差学习和感知损失的图像去噪", 中国图象图形学报, no. 10 * |
| 郭恒意;贾振堂;: "结合残差密集块的卷积神经网络图像去噪方法", 计算机工程与设计, no. 07 * |
| 魏欣;郑玉甫;: "基于总分式密集连接网络的图像超分辨重建", 兰州交通大学学报, no. 06 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113379638A (en) * | 2021-06-23 | 2021-09-10 | 西安理工大学 | SAR image despeckle method based on wavelet transform and interval dense network |
| CN114129171A (en) * | 2021-12-01 | 2022-03-04 | 山东省人工智能研究院 | Electrocardiosignal noise reduction method based on improved residual error dense network |
| CN114638829A (en) * | 2022-05-18 | 2022-06-17 | 安徽数智建造研究院有限公司 | Anti-interference training method of tunnel lining detection model and tunnel lining detection method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112991199B (en) | 2024-02-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109035149B (en) | License plate image motion blur removing method based on deep learning | |
| CN112991199B (en) | Image high-low frequency decomposition noise removal method based on residual dense network | |
| CN111161360B (en) | Image defogging method of end-to-end network based on Retinex theory | |
| CN107133923B (en) | Fuzzy image non-blind deblurring method based on adaptive gradient sparse model | |
| CN108765330B (en) | Image denoising method and device based on global and local prior joint constraint | |
| CN111861894A (en) | Image motion blur removing method based on generating type countermeasure network | |
| Onur | Improved image denoising using wavelet edge detection based on Otsu’s thresholding | |
| CN114723630A (en) | Image deblurring method and system based on cavity double-residual multi-scale depth network | |
| Pires et al. | Image denoising using attention-residual convolutional neural networks | |
| CN112801899B (en) | Internal and external circulation driving image blind deblurring method and device based on complementary structure perception | |
| CN111047537A (en) | System for recovering details in image denoising | |
| Yang et al. | Mixed noise removal by residual learning of deep cnn | |
| CN112381725B (en) | Image restoration method and device based on depth convolution countermeasure generation network | |
| Thanh et al. | Distorted image reconstruction method with trimmed median | |
| Karthikeyan et al. | Hybrid approach of efficient decision-based algorithm and fuzzy logic for the removal of high density salt and pepper noise in images | |
| CN113379641A (en) | Single image rain removing method and system based on self-coding convolutional neural network | |
| AKINTOYE et al. | COMPOSITE MEDIAN WIENER FILTER BASED TECHNIQUE FOR IMAGE ENHANCEMENT. | |
| Wang et al. | A review of image denoising methods | |
| Latifi et al. | Image denoising using convolutional neural network | |
| Chen et al. | A pilot study on a new image enhancement method using spatial statistics | |
| Sargolzaei et al. | Impulse image noise reduction using fuzzy-cellular automata method | |
| Verma et al. | Analysis of Hybrid Wavelet Transformed Images to Improve Image Pixel Quality | |
| Padmapriya et al. | Blind Image Quality Assessment with Image Denoising: A Survey | |
| Charmouti et al. | Progression approach for image denoising | |
| CN113379643B (en) | Image denoising method based on NSST domain and Res2Net network |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |