CN116342406A - Ultra-high resolution image denoising method based on deep learning - Google Patents
Ultra-high resolution image denoising method based on deep learning Download PDFInfo
- Publication number
- CN116342406A CN116342406A CN202310119614.0A CN202310119614A CN116342406A CN 116342406 A CN116342406 A CN 116342406A CN 202310119614 A CN202310119614 A CN 202310119614A CN 116342406 A CN116342406 A CN 116342406A
- Authority
- CN
- China
- Prior art keywords
- image
- denoising
- ultra
- high resolution
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000013135 deep learning Methods 0.000 title claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 7
- 238000012360 testing method Methods 0.000 claims abstract description 6
- 238000010606 normalization Methods 0.000 claims abstract description 4
- 230000008569 process Effects 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000012545 processing Methods 0.000 abstract description 11
- 230000000694 effects Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 208000025174 PANDAS Diseases 0.000 description 2
- 208000021155 Paediatric autoimmune neuropsychiatric disorders associated with streptococcal infection Diseases 0.000 description 2
- 240000004718 Panda Species 0.000 description 2
- 235000016496 Panda oleosa Nutrition 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 239000002699 waste material Substances 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000000007 visual effect Effects 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- 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
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- 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)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a super-resolution image denoising method based on deep learning, and belongs to the technical field of image processing. The method comprises the following steps: step 1, constructing an ultrahigh resolution image cutting and encoding model; step 2, inputting the noise image and the clean image into a cutting and coding model to obtain an image block; step 3, carrying out normalization of different degrees on the image blocks, constructing an image block sequence, and obtaining training and testing data; step 4, constructing a denoising network model based on the image; step 5, inputting the image block into an image denoising network model to obtain a clean image block; and 6, decoding and splicing the clean image blocks output by the network into an original image to obtain the denoised ultrahigh resolution image. The method provided by the invention has good effect on the ultra-high resolution image, the network model can be trained end to end, and the practicability is higher in the actual ultra-high resolution image processing scene.
Description
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to an ultra-high resolution image denoising method based on deep learning.
Background
With the advancement of imaging technology, the resolution of images captured by cameras has increased greatly, and many ultra-high resolution (e.g., resolution in excess of 1 billion pixels) images have appeared, and the quality requirements for such images have increased. However, these ultra-high resolution images are generally much more noisy than ordinary images and do not meet the visual needs of people well.
Image denoising is a basic technology in the field of image processing, and due to the superiority of denoising performance, the current front-edge denoising method is generally based on deep learning. However, the biggest difficulty in processing ultra-high resolution images by deep learning is the problem of computational resources, and the existing denoising method based on deep learning generally directly inputs images into a network for training and reasoning to obtain denoised images, and the complicated convolution operations consume large computational resources, so that the input images cannot be oversized, and therefore, the ultra-large scale images cannot be directly applied to the denoising methods. One solution is to downsample such images to smaller sizes and then perform network processing, however this approach works poorly because a lot of information is lost during downsampling.
Disclosure of Invention
Aiming at the technical problem that the existing denoising method is mainly only aimed at common low-resolution images and cannot be directly used for ultrahigh-resolution images, the invention provides an ultrahigh-resolution image denoising method based on deep learning.
In order to solve the problems, the invention adopts the following technical scheme:
a denoising method of an ultra-high resolution image based on deep learning comprises the following steps:
step 1, obtaining an image with ultra-high resolution as a clean image, and obtaining a noise image by adding noise;
step 2, performing quadtree cutting and position coding on the clean image and the noise image respectively to obtain clean image blocks and noise image blocks with different sizes;
step 3, respectively normalizing the clean image blocks and the noise image blocks to different degrees to ensure that the sizes of all the image blocks are consistent, constructing a clean image block sequence and a noise image block sequence, and obtaining training and testing data;
step 4, constructing an image denoising network model, which comprises 3 stages for preliminary denoising, fine denoising and result fine tuning, wherein each stage comprises downsampling to enhance denoising performance;
step 5, inputting the noise image block sequence into the image denoising network model to obtain a clean image block;
and 6, decoding and splicing the clean image block obtained in the step 5 to obtain the denoised ultrahigh resolution image.
Further, when the step 2 performs quad tree cutting, the noise image is first regarded as a root node of a tree, whether the absolute value of the difference between the maximum value and the minimum value of all pixel values is smaller than a threshold value is judged, if so, the current node is considered to belong to a region with simple texture, the fourth is stopped, if so, the texture of the current node is considered to be still more complex, and the fourth is continued; the above process is repeated until all nodes reach the preset condition or the stop condition.
Further, the preset condition includes that the size of the current node reaches a minimum 128×128.
Further, the preset condition further includes a maximum size 2048×2048 of the current node, and if the size of the current node is greater than the maximum size when the current node meets the stop condition, the node is forced to be quartered.
Further, in the step 2, the position code takes the position of the first pixel on the left upper part of the node on the original ultrahigh resolution image, and meanwhile, the size of the node is calculated, and the position code on the noise figure is performed synchronously with the quadtree cutting.
In step 2, the noise image is cut in a quadtree and position-coded to obtain all image blocks, and then the image blocks at the corresponding positions are found on the clean image according to the position-coding, and the image blocks are cut and are in one-to-one correspondence with the image blocks of the noise image.
Further, in the step 3, the specific normalization method is as follows: downsampling image blocks greater than the minimum value to a minimum size such that all image blocks are the same size, all image blocks being arranged into a sequence of image blocks.
Further, in the step 4, the 3 stages are independent, the input and output sizes of the stages are the same, and the 3 stages are trained together.
Further, in the step 4, the number of downsampling decreases with the increase of the stages, then features are obtained through a plurality of convolution operations and activation functions, the features with low size are upsampled from low to high in sequence and fused to the features with high size, no information is lost in the upsampling process, and after a subsequent series of convolution operations, the output is connected as a residual.
Further, in step 6, the clean image blocks output by the image denoising network model are sequentially extracted, decoding is performed according to the encoded position information, the position is restored to the position on the original ultra-high resolution image, the size is restored to the size when the original image is cut, and all the image blocks are spliced to obtain the final denoised ultra-high resolution image.
According to the invention, the image is cut and preprocessed, the Convolutional Neural Network (CNN) is used for noise reduction after normalization, and finally the noise-reduced image is obtained, so that the deep learning method can be applied to the image, and a good noise removal result is obtained. The network model can be trained end to end, and has high practicability in an actual ultrahigh resolution image processing scene. In addition, the method of the invention can be used for denoising and can be popularized to various computer vision tasks.
Drawings
FIG. 1 is a basic flow chart of the method of the present invention;
FIG. 2 is a detailed view of an image denoising network model constructed according to the present invention;
FIG. 3 is an effect diagram implemented by the present invention, including an original noise diagram, a clean diagram after cutting, and an output denoising diagram;
fig. 4 shows the performance of the present invention compared to other prior art operations, including peak signal to noise ratio PSNR and structural similarity SSIM.
Detailed Description
In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the present invention in terms of techniques, compositions, and effects achieved. The present invention is not described in greater detail with respect to some of the common sense problems in the art. To prevent confusion with other examples, the present invention will only be described in terms of its own technology.
The invention aims to solve the problem that the existing denoising method is not suitable for an ultrahigh resolution image, and provides a denoising method for the ultrahigh resolution image, so that a deep learning method can be applied to the ultrahigh resolution image. The specific method flow is shown in fig. 1, and comprises the following steps:
step 1, constructing an ultrahigh resolution image cutting and encoding model, which comprises image multi-stage cutting and position encoding based on a quadtree;
step 2, inputting the noise image and the clean image into a cutting and coding model respectively to obtain an image block;
step 3, normalizing the image blocks to different degrees to enable the sizes of all the image blocks to be consistent, constructing an image block sequence, and obtaining training and testing data;
step 4, constructing an image denoising network model, which comprises 3 stages, wherein the stages respectively realize preliminary denoising, fine denoising and result fine tuning, and each stage comprises downsampling to enhance denoising performance;
step 5, inputting the image block sequence into an image denoising network model to obtain a clean image block;
and 6, decoding and splicing the clean image blocks output by the network to obtain the denoised ultrahigh resolution image.
The invention is based on deep learning, so that it is necessary to construct an image dataset for training or the like before processing takes place. The invention uses the billion pixel dataset PANDA dataset disclosed in academia as clean images that contain ultra-high resolution images of a large number of scenes, each with a resolution between one hundred million and billion. The present embodiment adds different levels of additive gaussian noise conforming to normal distribution as noise images thereon, with noise standard deviations of three levels of 10, 30, 50.
In this embodiment, the image is cut first, so that the purpose of not losing information is achieved. One simple method is to make carpet cuts directly and then input into the network for processing, however, this method is slow. In addition, such images have some specificities, for example, the images are taken of large outdoor scenes with wide viewing angles, such as cities, starry sky, earth and the like, the images contain a large number of non-textured areas, the denoising of the areas is very convenient, and complicated analysis is required for some areas with complicated textures, so that if carpet cutting is performed on the areas with different complexity and the same processing is applied, further waste in time is brought, and the large-scale data volume is particularly obvious. Therefore, the present embodiment proposes a quadtree-based image multi-stage cutting method, which aims to avoid the waste of computing resources by performing different degrees of processing by using the complexity differences of different areas of the image itself while performing rapid cutting.
In step 1, the noise image is regarded as a root node of a tree, the noise image is continuously divided into four parts, the complexity of the node is required to be calculated before each four parts, and the complexity of the node is judged specifically by judging the difference between the maximum value and the minimum value in the node. By setting a threshold k, the range of the value can be adjusted according to the noise level of the image, and by experiments, when the noise standard deviation is 10, 30 and 50 under the 8bit image, the value of k is set to be about 25, 50 and 75, so that the final denoising result can be optimized. In the process of the fourth score, whether the complexity of the node is smaller than a threshold value is continuously judged, if so, the node is considered to belong to the area with simple texture, the fourth score is stopped, and if so, the texture of the node is considered to be still more complex, and the fourth score is continued. This process is repeated until all nodes have reached the minimum size or stop condition specified by the present invention. For the minimum size, the segmentation result is not ideal, so that the final denoising performance is reduced, the number of image blocks is increased, the appearance of the boundary of the image blocks is reduced, and the final denoising result is optimal by using the minimum size of 128 x 128 through experiments. In addition, in order to prevent the image block cut out on some images with less texture information from being excessively large, a maximum size 2048×2048 is added, i.e., if the node size is larger than the maximum size when the stop condition is satisfied, the node is forced to be quarter-divided. In the process, each node is subjected to position coding, namely the size of each node and the position of the first pixel on the left upper side of each node on the original ultra-high resolution noise diagram are recorded, so that subsequent output decoding is facilitated to restore the original diagram. This process can be expressed mathematically as follows:
where x is a node, x 1 ,x 2 ,x 3 ,x 4 For four child nodes, QD (·) represents a quarter cut, and f (·) is the complexity of the decision node, i.e., |max (x) -min (x) |.
In order to obtain image blocks corresponding to the noise map and the clean map one by one, the embodiment firstly performs the cutting step on the noise map, and after all the image blocks are obtained, the image blocks with the same position and size are directly cut on the clean map according to the position and size information of all the nodes, so that all the two types of image blocks can be in one-to-one correspondence.
For step 3, after simple and complex image blocks are obtained, the simpler the image blocks the more powerful noise reduction process can be applied. Specifically, complex image blocks are temporarily not processed, and simple image blocks are downsampled to different extents. That is, for image blocks larger than 128×128, the size is uniformly reduced to 128×128, the larger the size is, the simpler the texture is, the larger the downsampling force is, and the more direct the noise reduction means is. These image blocks are then arranged into a sequence, and finally, an image sequence of 128 x 128 sizes is obtained. This is done on both the noise and clean maps. This process can be expressed mathematically as follows:
x′＝Down(x)
where x is the node of a simple image block, x' is the node after downsampling, and Down (·) is the downsampling process.
For step 4, in order to obtain an efficient denoising model, the embodiment constructs a progressive denoising convolutional neural network model based on multiple stages and multiple scales, specifically including three stages, if the stages are set too little, the performance is reduced, too much improvement is not brought, but the computational complexity is increased. In one stage, the size of the input noise image (b, c, h, w) is (1,3,128,128), b represents the batch size of batch_size, c represents the channel number, h represents the height image block height, w represents the width of the width image block, the input noise image is downsampled for multiple times, downsampling is performed for 3 times in one stage, downsampling is performed for 2 times in two stages, three-stage is performed for 1 time, each downsampling reduces the height and width to one half of the original image, the corresponding dimension c is increased by 4 times, thus obtaining three downsampled inputs (1,12,64,64), (1,48,32,32), (1,192,16,16), the input is subjected to convolution operation to obtain a feature image, the convolution kernel size is 3*3, the channel number of the feature image is increased along with the reduction of the height and width, after the multiple convolution operation is performed, the feature image with the low size is upsampled, the dimension c is reduced by 1/4 in the original downsampled inverse process, the dimension c is increased by 2 times in the original dimension, the feature image with the previous dimension c is subjected to the subsequent convolution operation, and the feature image with the specific dimension c is subjected to the subsequent convolution operation. This process is repeated until the low-size feature maps are fused to the original size. After a series of convolutions, the output channel c at the last layer is 3, so as to ensure that the output is the same as the original input in size. To guarantee the validity of this phase, the residual connection from input to output is added, i.e. expressed mathematically as:
x 1 x 0 +F(x 0 )
wherein x is 0 For input, x 1 F (·) is the output of a stage, and F (·) is all operations of a stage.
The following two and three stages do similar operations, and the operation at the back end is more fine-tuning of the output, so the number of corresponding downsampling is reduced. ReLU is added as an activation function after convolution operation, and the whole network structure is shown in figure 2.
For step 5, the invention uses noise and clean image sequences as training data, uses Adam optimizer, loss function is MSE Loss, sets batch_size to 8, 300 epochs, iterates 153600 times, learning rate is 0.0001 for the first 100 epochs, then 100 epochs are reduced to 0.00001, finally 100 epochs are reduced to 0.000001, and the GPU is single RTX 3090. After the denoising model is trained, testing is carried out on test data, and only a noise image block sequence is input to obtain a denoised image block sequence.
And step 6, the obtained denoising image blocks are spliced one by one, and the specific method is that the denoising image blocks are decoded according to the encoded position information, the positions are restored to the positions on the original ultra-high resolution image, the sizes of the denoising image blocks are restored to the sizes of the original image which are not cut, and the method for restoring the sizes is an image interpolation upsampling method. Finally, all the image blocks are spliced to obtain the denoised ultrahigh resolution image.
The effect of the method used in the present invention on denoising is shown in fig. 3. To illustrate the advantages of the present invention, the peak signal to noise ratio PSNR versus structural similarity SSIM performance of the different methods on the Panda dataset, which is the disclosed hundred million-level pixel image dataset, dnCNN is the most classical deep learning denoising method, ffdnaet is an earlier proposed denoising method over DnCNN, cbdnaet is an improvement of ffdnaet, SGN is a very good performing denoising method, and the final result is shown in fig. 4.
Claims (10)
1. The ultra-high resolution image denoising method based on deep learning is characterized by comprising the following steps of:
step 1, obtaining an image with ultra-high resolution as a clean image, and obtaining a noise image by adding noise;
step 2, performing quadtree cutting and position coding on the clean image and the noise image respectively to obtain clean image blocks and noise image blocks with different sizes;
step 3, respectively normalizing the clean image blocks and the noise image blocks to different degrees to ensure that the sizes of all the image blocks are consistent, constructing a clean image block sequence and a noise image block sequence, and obtaining training and testing data;
step 4, constructing an image denoising network model, which comprises 3 stages for preliminary denoising, fine denoising and result fine tuning, wherein each stage comprises downsampling to enhance denoising performance;
step 5, inputting the noise image block sequence into the image denoising network model to obtain a clean image block;
and 6, decoding and splicing the clean image block obtained in the step 5 to obtain the denoised ultrahigh resolution image.
2. The ultra-high resolution image denoising method based on deep learning according to claim 1, wherein when the step 2 performs quadtree cutting, the noisy image is regarded as a root node of a tree, whether the absolute value of the difference between the maximum value and the minimum value of all pixel values is smaller than a threshold value is judged, if so, the current node is considered to be already in a region with simple texture, the fourth is stopped, if so, the texture of the current node is considered to be still more complex, and the fourth is continued; the above process is repeated until all nodes reach the preset condition or the stop condition.
3. The method for denoising an ultrahigh resolution image based on deep learning according to claim 2, wherein the preset condition comprises that the size of the current node reaches 128×128.
4. A deep learning-based ultra-high resolution image denoising method according to claim 3, wherein the preset condition further comprises a maximum size 2048×2048 of the current node, and if the current node satisfies the stop condition, the node is forced to be quarter-sized if the node size is larger than the maximum size.
5. The method for denoising ultra-high resolution image based on deep learning according to claim 1, wherein in the step 2, the position coding is performed by taking the position of the first pixel on the left upper part of the node on the original ultra-high resolution image, calculating the size of the node, and performing the position coding on the noise figure in synchronization with the quadtree cutting.
6. The method for denoising ultra-high resolution image based on deep learning according to claim 1, wherein in step 2, the noise image is subjected to quadtree cutting and position coding to obtain all image blocks, and then the image blocks at the corresponding positions are found on the clean image according to the position coding, and are cut and are in one-to-one correspondence with the image blocks of the noise image.
7. The ultra-high resolution image denoising method based on deep learning according to claim 1, wherein in the step 3, the specific normalization method is as follows: downsampling image blocks greater than the minimum value to a minimum size such that all image blocks are the same size, all image blocks being arranged into a sequence of image blocks.
8. The ultra-high resolution image denoising method based on deep learning according to claim 1, wherein in the step 4, 3 stages are independent of each other, the input and output dimensions of each stage are the same, and the 3 stages are trained together.
9. The method according to claim 8, wherein in the step 4, the number of downsampling is reduced with the increase of the stage, then features are obtained through a plurality of convolution operations and activation functions, the features with low size are upsampled from low to high in sequence and fused to the features with high size, no information is lost in the upsampling process, and the output is used as a residual connection after a subsequent series of convolution operations.
10. The method for denoising ultra-high resolution image based on deep learning according to claim 1, wherein in step 6, the clean image blocks outputted by the image denoising network model are sequentially extracted, decoded according to the encoded position information thereof, the position thereof is restored to the position on the original ultra-high resolution image, the size thereof is restored to the size when cut on the original image, and all the image blocks are spliced to obtain the final denoised ultra-high resolution image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310119614.0A CN116342406A (en) | 2023-02-16 | 2023-02-16 | Ultra-high resolution image denoising method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310119614.0A CN116342406A (en) | 2023-02-16 | 2023-02-16 | Ultra-high resolution image denoising method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116342406A true CN116342406A (en) | 2023-06-27 |
Family
ID=86883019
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310119614.0A Pending CN116342406A (en) | 2023-02-16 | 2023-02-16 | Ultra-high resolution image denoising method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116342406A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116757966A (en) * | 2023-08-17 | 2023-09-15 | 中科方寸知微(南京)科技有限公司 | Image enhancement method and system based on multi-level curvature supervision |
| CN117423113A (en) * | 2023-12-18 | 2024-01-19 | 青岛华正信息技术股份有限公司 | Adaptive denoising method for archive OCR (optical character recognition) image |
-
2023
- 2023-02-16 CN CN202310119614.0A patent/CN116342406A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116757966A (en) * | 2023-08-17 | 2023-09-15 | 中科方寸知微(南京)科技有限公司 | Image enhancement method and system based on multi-level curvature supervision |
| CN117423113A (en) * | 2023-12-18 | 2024-01-19 | 青岛华正信息技术股份有限公司 | Adaptive denoising method for archive OCR (optical character recognition) image |
| CN117423113B (en) * | 2023-12-18 | 2024-03-05 | 青岛华正信息技术股份有限公司 | Adaptive denoising method for archive OCR (optical character recognition) image |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110120011B (en) | Video super-resolution method based on convolutional neural network and mixed resolution | |
| CN111028150B (en) | Rapid space-time residual attention video super-resolution reconstruction method | |
| CN108765296B (en) | Image super-resolution reconstruction method based on recursive residual attention network | |
| CN116342406A (en) | Ultra-high resolution image denoising method based on deep learning | |
| CN110738605A (en) | Image denoising method, system, device and medium based on transfer learning | |
| CN105825477B (en) | The Remote sensed image super-resolution reconstruction method merged based on more dictionary learnings with non-local information | |
| CN112257766B (en) | Shadow recognition detection method in natural scene based on frequency domain filtering processing | |
| CN109785236B (en) | Image super-resolution method based on super-pixel and convolutional neural network | |
| CN110751597A (en) | Video super-resolution method based on coding damage repair | |
| CN113129212B (en) | Image super-resolution reconstruction method and device, terminal device and storage medium | |
| CN114581304B (en) | Image super-resolution and defogging fusion method and system based on circulation network | |
| CN111696033A (en) | Real image super-resolution model and method for learning cascaded hourglass network structure based on angular point guide | |
| CN113554058A (en) | Method, system, device and storage medium for enhancing resolution of visual target image | |
| CN105590296B (en) | A kind of single-frame images Super-Resolution method based on doubledictionary study | |
| CN103400346B (en) | The video super-resolution method of autoregression model is guided based on adaptive super-pixel | |
| CN110264483B (en) | Semantic image segmentation method based on deep learning | |
| CN113947538A (en) | Multi-scale efficient convolution self-attention single image rain removing method | |
| CN113962882A (en) | JPEG image compression artifact eliminating method based on controllable pyramid wavelet network | |
| CN103400394A (en) | Video super-resolution method based on feature-oriented variational optical flow | |
| CN112529801B (en) | Motion blur restoration method based on high-frequency image block estimation blur kernel | |
| CN111626945B (en) | Depth image restoration method based on pixel-level self-similar model | |
| CN104123707A (en) | Local rank priori based single-image super-resolution reconstruction method | |
| CN112907456B (en) | Deep neural network image denoising method based on global smooth constraint prior model | |
| CN113077385A (en) | Video super-resolution method and system based on countermeasure generation network and edge enhancement | |
| Chang et al. | A genetic algorithm approach to image sequence interpolation |
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 |