CN109903223B - Image super-resolution method based on dense connection network and generation type countermeasure network - Google Patents
Image super-resolution method based on dense connection network and generation type countermeasure network Download PDFInfo
- Publication number
- CN109903223B CN109903223B CN201910033048.5A CN201910033048A CN109903223B CN 109903223 B CN109903223 B CN 109903223B CN 201910033048 A CN201910033048 A CN 201910033048A CN 109903223 B CN109903223 B CN 109903223B
- Authority
- CN
- China
- Prior art keywords
- image
- resolution
- network
- resolution image
- loss
- 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.)
- Active
Links
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention provides an image super-resolution method based on a dense connection network and a generation type countermeasure network, namely, a low-resolution image is converted into high resolution. The method includes three models: a dense connectivity network based generator, a VGG19 network based feature extractor; convolutional neural network discriminators based on super-resolution images. In this method, the input of the generator is a low resolution image, and a high resolution image is generated. The feature extractor is used for extracting the content features of the generated high-resolution image, and the discriminator is used for discriminating whether the distribution information of the generated high-resolution image is accurately expressed. And (3) performing joint optimization on the three models. The optimization process uses the mean square error loss between pixels, the VGG-19 based content loss, and the contrast loss at the arbiter. The invention generates the high resolution image according to the low resolution image, and is suitable for image data of various scenes.
Description
Technical Field
The invention belongs to the technical field of image super-resolution, and particularly relates to a design based on a dense connection network structure and an image super-resolution method of a generated type countermeasure network, which is suitable for super-resolution of 4 times of images.
Background
Image super-resolution is one of the classical tasks in computer vision, with the goal of reconstructing a high resolution image from a low resolution input image. It is commonly used in some particular important fields, such as license plate images and image super-resolution of medical images. However, image super resolution is an irreversible problem, since there are several possible original values for any low resolution pixel. To solve this problem, the model requires a powerful mapping function from low resolution image to high resolution image and is to be learned based on sufficient a priori knowledge.
Early image super-resolution methods used bicubic interpolation, which, although fast and efficient, resulted in blurring of the generated image. In recent years, super resolution has been mainly two ways, one of which is an internal-based example method, which uses self-similarity properties and generates sub-image blocks from an input-possible image. On the other hand, the external instance-based approach learns the mapping function of the training dataset from low resolution to high resolution. In recent years, methods based on Convolutional Neural Networks (CNNs) have been proposed in the academy, which belong to the second approach, wherein the input is a bicubic interpolated image sub-block. This approach can potentially learn powerful mapping functions. Such super-resolution convolutional neural networks (SRCNN) can produce very clear generated graphs compared to other machine learning methods. Better super-resolution results are produced as the number of layers of the network is progressively deeper. The appearance of dense connection network can make the characteristic transmission effect better, and can greatly increase the network depth, this can make the network obtain bigger receptive field, obtains better function mapping effect.
The Generated Antagonism Network (GANs) is a deep learning model, and is one of the most promising methods for unsupervised learning on complex distribution in recent years. The model is built up of (at least) two modules in a frame: a generator and a arbiter that learn in game with each other can produce a fairly good output. GAN provides a sophisticated model in which the arbiter measures the distribution between the generated output and the target data. And training the generator based on the difference between its discrimination and the generated data. WGAN (Wasserstein GAN) provides a more stable training framework. Bulldozer distance (EMD) is used during the discriminant process rather than the previous Jensen-Shannon divergence to better measure the distribution between different data. In recent years, deep learning has achieved good experimental and application effects in related fields such as computer vision. The objective of deep learning is to learn robust semantic information with rich expression capability from complex data, however, the image generated by deep learning is based on a mean square loss function, which often brings about some blurring effect, so that the combination of the generation type countermeasure network and the image super-resolution becomes the key for solving the problem.
Christian Ledig et al, paper "Photo-Realistic Single Image Super-Resolution Usinga Generative Adversarial Network" and found that over 4 times the super-resolution of the image would blur the image smoothly, analyzing the Mean Square Error (MSE) loss function between the super-resolution image and the high resolution image pixels did not update the generator well. The idea of introducing content loss based on VGG-19 network and counterloss of a discriminator is to discriminate the original image and the generated image so that the super-resolution image is similar to the original image in style. Although the paper also adopts a generating type antagonism network to train the generator and introduces multiple losses to increase the sense of reality of the image, the method uses a network with fewer layers and is not robust, and a discriminator adopts a Jensen-Shannon divergence which cannot well measure the true distance between the super-resolution image and the high-resolution image. The present invention improves on these two points, namely using a deeper dense connection network as the generator and EMD for the arbiter.
Disclosure of Invention
The technical problems to be solved by the invention are as follows: the image super-resolution method based on the dense connection network and the generated type countermeasure network overcomes the defect of blurring of the existing super-resolution image, and has the advantages that the depth of the network is deepened through the dense connection network to generate a better super-resolution effect, and meanwhile, the image super-resolution effect is enabled to be more consistent with a real effect and clearer through EMD.
The technical proposal of the invention is as follows: an image super-resolution method based on a dense connection network and a generated countermeasure network provides a novel image super-resolution network architecture of the dense connection network, the network architecture is applied to a mechanism of the generated countermeasure network, the problems of mode collapse and large training errors of the traditional generated countermeasure network are solved, and an EMD is used for distinguishing the distance between a generated image and a tag high-resolution image.
The specific implementation method comprises the following steps:
firstly, constructing training data; randomly cropping the image of the image dataset into a square area with the size of 120×120 pixels, reducing the cropped image to the original size of 1/4 (30×30 pixels) by using bicubic interpolation method, generating a low-resolution training image, and using the 120×120 high-resolution image before scaling as tag data;
in a second step, the training network builds a dense connection network for the generator, inputting the low resolution training images generated for the first step, the images generated being referred to as super resolution images. Training a convolutional neural network to act as a discriminator for discriminating that the generated super-resolution image does not belong to the high-resolution image described in the first step. The training generator then makes it generate super-resolution images which can fool the arbiter, and makes the arbiter misinterpret the images generated by the generator as the high-resolution images described in the first step. Training the two alternately to a certain number of times to obtain a final dense connection network;
and thirdly, generating a high-resolution image, namely directly inputting the low-resolution image to be detected into a final dense connection network in the second step, wherein an output image generated by the network is the target high-resolution image.
Constructing a dense generation type network model in the second step, constructing a dense connection network for a generator, adopting a depth structure to ensure the data fitting capability of the network, adopting a jump dense connection to ensure the transmission of image features of different levels in order to more easily train the network to alleviate the problem of gradient disappearance, wherein the whole network model totally comprises 4 jump dense modules, each dense module respectively comprises 6, 12, 24, 16 convolution layers of 3×3 and 1×1 convolution layers, each convolution layer can generate different kinds of feature images, the feature images are input to the next convolution layer to generate more complex feature images, and the network model adopts a linear correction unit (Relu) as an activation function; the second last layer and the third layer are deconvolution layers used for amplifying the feature images, and the final layer is a reconstruction layer which can restore the feature images generated by the previous convolution operation into images. The convolution layer pairs inside the skip-dense module all take the following formula operation, the first layer taking the signature x generated from the 1 to l-1 convolution layers 0 ，x 1 ...，x l-1 As input, its output x l The following formula is shown:
x l ＝H l ([x 0 ，x 1 ，...，x l-1 ])
wherein [ x ] 0 ，x 1 ，...，x l-1 ]Representing feature map stitching operations, each dense module is followed by a conversion layer for feature map quantity compression and more complexityAnd (5) feature mapping. H l Is the operation of the first convolution layer.
The network combines three loss functions to enable the super-resolution image effect to be clearer, calculate the MSE loss of the tag image and the generated image, and set I LR For low resolution images, I HR For the original high resolution image, G is a dense mapping operation that is a networked model, and the MSE loss function can then be calculated as follows:
w, H is the length and width of the tag high resolution image, C is the number of image channels, and the invention is set to 3. In addition to MSE loss, the content loss on VGG-19 of the generated image and the label image needs to be calculated for comparing semantic information of the super-resolution image and the high-resolution image:
wherein, A and B are the length and width of the layer characteristic diagram and N is the number of channels of the characteristic diagram, which are mapping functions of the 37 th layer based on a pretrained VGG-19. In addition, the EMD distance loss of the generated image and the label image on the discriminator is calculated. The loss function is as follows:
l adv ＝-D(G(I LR ))
g is a generator, D is a direct distance between the discriminator and the generated image for comparing the tag high resolution image. The three loss functions are then combined according to the following specific gravity to obtain a total loss function l total . Updating parameters of the densely connected network by adopting a back propagation algorithm:
l total ＝l mse +0.07l content +0.01l adv
compared with the prior art, the invention has the advantages that: the invention uses a pair of low resolution image and high resolution image pairs to train the network, uses a mean square loss function (MSE) to minimize the pixel error of the super resolution image and the high resolution image, uses VGG-19 network as the comparison loss function of the super resolution image and the high resolution image characteristics, can make the semantic characteristics of the two be as close as possible, and uses EMD as the distance between the two of the discriminators to make the two be as close as possible in the image style.
Drawings
FIG. 1 is a block diagram of a generated countermeasure network of the present invention;
fig. 2 is a dense connectivity network (generator) training schematic of the present invention.
Detailed Description
The following description of the embodiments of the invention is presented in conjunction with the accompanying drawings to provide a better understanding of the invention to those skilled in the art.
The invention relates to an image super-resolution method based on a dense connection network and a generation type countermeasure network, which belongs to the field of image super-resolution, and comprises the steps of firstly constructing training data, randomly cutting an image of an image dataset into square areas with fixed sizes, reducing the cut image by fixed times by using a bicubic interpolation method, generating a small image to serve as the training data, and using an original cut image as tag data. Then a network model is constructed, the input of the model is an original small image, and the output is an image after super resolution. And dense connections are all taken within each dense module, with residual connections taken by the external modules. And uses the look-ahead modification unit as an activation function. The super-resolution method can generate clear images and is suitable for various types of image data.
As shown in fig. 1 and 2, the invention comprises the following steps:
1. constructing training data, randomly clipping an image of an image dataset into a square area with a fixed size, and reducing the clipped image by a fixed multiple by using bicubic interpolation to generate a small image serving as the training data and an original clipping image serving as the tag data.
2. The contrast training dense connection network is used for inputting image pairs in training data into a network model to generate super-resolution images, then inputting high-resolution images and low-resolution images into a discriminator, training the discriminator to accurately discriminate the two images, fixing a discriminator training generator after training the discriminator, calculating the MSE loss of a label image and a generated image, calculating the feature loss of the generated image and the label image on VGG-19 and calculating the EMD loss of the generated image and the label image on the discriminator according to the mode of figure 2. The parameters of the dense connectivity network are then updated using a back propagation algorithm.
3. And selecting any image from the test set, inputting the selected image into a dense connection network, and obtaining a super-resolution image which is a super-resolution high-resolution image.
Compared with the prior art, the invention has obvious advantages and beneficial effects: the invention not only uses dense connection network to deepen the network depth of super-resolution so as to make the effect better, but also adopts the training thought of WGAN so as to remove the blurring of the super-resolution image, which is more true.
The specific implementation steps are as follows:
(1) The database in the implementation process of the invention is derived from a public standard data set ImageNet, and 60000 images are randomly selected to be used as training data. During the training process, images of 120×120 size are randomly cropped and then reduced by 4 times to produce a low resolution training image of 30×30 size. While the pre-scaled image serves as contrast high resolution label data.
(2) As shown in FIG. 2, a dense connection network is constructed for a generator, the invention adopts a depth structure to ensure the data fitting capability of the network, and adopts a jump dense connection to ensure the transmission of image features of different levels in order to more easily train the network to alleviate the problem of gradient disappearance, the whole network model totally comprises 4 jump dense modules, each dense module respectively comprises 6, 12, 24 and 16 convolution layers with the size of 1 multiplied by 3, and each convolution layer generates different types of special characters after each convolution layerThe feature maps are input to the next convolution layer to generate more complex feature maps, and the model adopts a linear correction unit (Relu) as an activation function; the penultimate and third layers are deconvolution layers used for amplifying the characteristic images, and the final layer is a reconstruction layer which can restore the characteristic images generated by the previous convolution layer into images. The convolution layer pairs inside the skip-dense module all take the following formula operation, the first layer taking the signature x generated from the 1 to l-1 convolution layers 0 ，x 1 …，x l-1 As an input, the following formula is shown:
x l ＝H l ([x 0 ，x 1 ，...，x l-1 ]) (1)
wherein [ x ] 0 ，x 1 ，...，x l-1 ]Representing a feature map stitching operation, each dense module is followed by a translation layer for feature map number compression and more complex feature mapping.
(3) The image pairs in the training data are input into a network model, and the network is trained by adopting a packet gradient descent method, namely, 32 image pairs are input for each training iteration. Firstly, a discriminator needs to be adjusted to accurately distinguish a super-resolution image from a high-resolution image, and the discriminator needs to judge the high-resolution image and judge the super-resolution image as much as possible, so that the loss is as follows:
max {||D||≤1} (E[D(I HR )]-E[D(I SR )] (2)
wherein E represents the desire, I SR ，I HR Representing the super-resolution image and the high-resolution image, wherein D is less than or equal to 1, and the D is required to satisfy a 1-Lipschitz function, namely, the difference between the high-resolution image and the super-resolution image discriminated by the discriminator cannot exceed the linear increase with the slope of 1.
After training the discriminator, training a generator based on a dense connection network, calculating the MSE loss of a label image and a generated image, and setting I LR For low resolution images, I HR For the original high resolution image, G is a dense mapping operation that is a networked model, and the MSE loss function can then be calculated as follows:
w, H is the length and width of the tag high resolution image, C is the number of image channels, and the invention is set to 3.m is the number of samples in a group. In addition to this, the content loss of the generated image and the label image on VGG-19 needs to be calculated:
wherein, A and B are the length and width of the layer characteristic diagram and N is the number of channels of the characteristic diagram, which are mapping functions of the 37 th layer based on a pretrained VGG-19. m is the number of training samples in a group. Phi () is a functional representation corresponding to the feature extraction network based on VGG-19, besides, the invention also needs to calculate the EMD distance loss of the generated image and the label image on the discriminator. The loss function is as follows:
g is a generator, D is a direct distance between the discriminator and the generated image for comparing the tag high resolution image.
The three above-mentioned loss functions are combined according to the following specific gravity to obtain a total loss function l total . This ratio is the combination found to be most effective in experiments. And updating parameters for adjusting the dense connection network using a back propagation algorithm.
l total ＝l mse +0.07l content +0.01l adv (6)
(4) The invention is finally used for the application to connect the network densely, a group of images are selected from the test set and input into the network, and the images generated by the network are final high-resolution images.
Claims (5)
1. An image super-resolution method based on a dense connection network and a generated type countermeasure network is characterized in that: the method comprises the following four steps:
the method comprises the steps of firstly, downsampling a tag high-resolution image into a low-resolution image, inputting the low-resolution image into a generator, and generating a high-resolution image; calculating the mean square error loss between pixels of the generated high-resolution image and the label high-resolution image;
the second step, the characteristic extractor extracts the characteristic content of the generated high-resolution image and the label high-resolution image, and calculates the content loss between the corresponding characteristic images;
thirdly, scoring the generated high-resolution image by the discriminator, wherein the higher the score is, the closer the generated image distribution is to the label high-resolution image and the countermeasure loss based on the discriminator is generated;
fourth, three loss functions are combined according to a certain proportion, and generator parameters are optimized;
in the first step: the generator has the following characteristics:
(1) Adopting a dense connection network structure, wherein the depth of the dense connection network structure exceeds 100 layers;
(2) The system comprises 4 jump dense modules, wherein each dense module is provided with a conversion layer;
(3) Upsampling using a deconvolution layer;
(4) Reconstructing an RGB color image using a convolutional layer with an input channel of 256 and an output channel of 3;
when a dense connection network is constructed for a generator, a depth structure is adopted to ensure the data fitting capability of the network, meanwhile, in order to more easily train the network to alleviate the problem of gradient disappearance, a jump dense connection is adopted to ensure the transmission of image features of different levels, a network model of the dense connection network totally comprises 4 jump dense modules, and a conversion layer is arranged behind each dense module and used for compressing the number of feature images and mapping more complex features; each dense module contains 6, 12, 24, 16 convolution layers of 3×3 and convolution layers of 1×1, and different kinds of characteristic diagrams are generated after each convolution layer, and the characteristic diagrams are input to the next convolution layer to generate more complex characteristic diagrams; the dense connection network adopts a linear correction unit Relu as an activation function; the penultimate and third layers are deconvolution layers used for amplifying the feature images, and the last layer is a reconstruction layer which restores the amplified feature images into images of RGB channels.
2. The image super-resolution method based on dense connection network and generation type countermeasure network according to claim 1, wherein: the first step is realized as follows:
(1) Randomly cropping the collected high-resolution images into a square area with the size of 120 multiplied by 120;
(2) The clipped image was scaled down to 0.25 to 30×30 times the original image using bicubic interpolation as a low resolution training image, and 120×120 before scaling was used as a new high resolution image label high resolution image.
3. The image super-resolution method based on dense connection network and generation type countermeasure network according to claim 1, wherein: the convolution layers inside the jump dense module all adopt the operation of the following formula, and the first layer is used for generating a characteristic diagram x from 1 to l-1 convolution layers 0 ,x 1 …,x l-1 As input, its output x l Expressed as the following formula:
x l ＝H l ([x 0 ,x 1 ,…,x l-1 ])
wherein [ x ] 0 ,x 1 ,…,x l-1 ]Representing matrix stitching operations, H l Is the operation of the first convolution layer.
4. The image super-resolution method based on dense connection network and generation type countermeasure network according to claim 1, wherein: in optimizing the generator, three loss functions need to be calculated,
(1) First, calculate the Mean Square Error (MSE) loss of the generated high resolution image and the tag high resolution image, sum the squares of the differences between each pixel and average the differences by comparing the twoMean Square Error (MSE) loss l mse The formula is as follows:
I LR to input 30×30 low resolution images for training, I HR For the tag high-resolution image, G is a generator, 120×120 high-resolution images can be generated through 30×30 low-resolution images, W and H are the length and width of the tag high-resolution image, and C is the number of color channels for generating the high-resolution image;
(2) The second loss function is content loss, and the feature loss of the generated high-resolution image and tag image on the VGG19 is calculated as follows: l (L) content
Wherein, A and B are the length and width of the 37 th layer characteristic diagram, N is the number of channels of the characteristic diagram, and phi is the function representation corresponding to the characteristic extraction network based on VGG-19;
(3) The third loss function is a counterloss, which is used to calculate the EMD distance-based loss of the generated image and the label image on the arbiter, expressed as:
l adv ＝-D(G(I LR ))
g is a generator, and D is the distance between the label high-resolution image and the generated high-resolution image distribution.
5. The image super-resolution method based on dense connection network and generative countermeasure network of claim 4, wherein: the three loss functions are synthesized according to the set proportion to obtain a total loss function l total ，l total ＝l mse +λ 1 l content +λ 2 l adv The method comprises the steps of carrying out a first treatment on the surface of the Updating l total When adopting the reverse directionPropagation algorithm adjusts parameters of dense connection network, lambda 1 And lambda (lambda) 2 The super parameters are set by specific experiments.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910033048.5A CN109903223B (en) | 2019-01-14 | 2019-01-14 | Image super-resolution method based on dense connection network and generation type countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910033048.5A CN109903223B (en) | 2019-01-14 | 2019-01-14 | Image super-resolution method based on dense connection network and generation type countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109903223A CN109903223A (en) | 2019-06-18 |
| CN109903223B true CN109903223B (en) | 2023-08-25 |
Family
ID=66943855
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910033048.5A Active CN109903223B (en) | 2019-01-14 | 2019-01-14 | Image super-resolution method based on dense connection network and generation type countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109903223B (en) |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110322528B (en) * | 2019-06-26 | 2021-05-14 | 浙江大学 | Nuclear magnetic resonance brain image vessel reconstruction method based on 3T and 7T |
| CN112188237A (en) * | 2019-07-04 | 2021-01-05 | 国家广播电视总局广播电视科学研究院 | Program distribution method and device, receiving method, terminal device and medium |
| CN110443867B (en) * | 2019-08-01 | 2022-06-10 | 太原科技大学 | CT image super-resolution reconstruction method based on generation countermeasure network |
| CN110717857A (en) * | 2019-09-29 | 2020-01-21 | 中国科学院长春光学精密机械与物理研究所 | Super-resolution image reconstruction method and device |
| CN110826567B (en) * | 2019-11-06 | 2023-04-07 | 北京字节跳动网络技术有限公司 | Optical character recognition method, device, equipment and storage medium |
| CN111080531B (en) * | 2020-01-10 | 2024-02-23 | 北京农业信息技术研究中心 | Super-resolution reconstruction method, system and device for underwater fish image |
| CN111476721B (en) * | 2020-03-10 | 2022-04-29 | 重庆邮电大学 | Wasserstein distance-based image rapid enhancement method |
| CN111402133A (en) * | 2020-03-13 | 2020-07-10 | 北京字节跳动网络技术有限公司 | Image processing method, image processing device, electronic equipment and computer readable medium |
| CN111476353B (en) * | 2020-04-07 | 2022-07-15 | 中国科学院重庆绿色智能技术研究院 | Super-resolution method of GAN image introducing significance |
| CN111583113A (en) * | 2020-04-30 | 2020-08-25 | 电子科技大学 | Infrared image super-resolution reconstruction method based on generation countermeasure network |
| CN111861870B (en) * | 2020-07-16 | 2022-03-15 | 南通大学 | End-to-end parallel generator network construction method for image translation |
| CN111861961B (en) * | 2020-07-25 | 2023-09-22 | 安徽理工大学 | Single image super-resolution multi-scale residual error fusion model and restoration method thereof |
| CN112102165B (en) * | 2020-08-18 | 2022-12-06 | 北京航空航天大学 | Light field image angular domain super-resolution system and method based on zero sample learning |
| CN116385270A (en) * | 2023-04-18 | 2023-07-04 | 华院计算技术(上海)股份有限公司 | Image-to-image method based on multiple loss and resolution |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107194872A (en) * | 2017-05-02 | 2017-09-22 | 武汉大学 | Remote sensed image super-resolution reconstruction method based on perception of content deep learning network |
| CN108765291A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | Super resolution ratio reconstruction method based on dense neural network and two-parameter loss function |
| CN108765290A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
| CN108765512A (en) * | 2018-05-30 | 2018-11-06 | 清华大学深圳研究生院 | A kind of confrontation image generating method based on multi-layer feature |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018053340A1 (en) * | 2016-09-15 | 2018-03-22 | Twitter, Inc. | Super resolution using a generative adversarial network |
-
2019
- 2019-01-14 CN CN201910033048.5A patent/CN109903223B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107194872A (en) * | 2017-05-02 | 2017-09-22 | 武汉大学 | Remote sensed image super-resolution reconstruction method based on perception of content deep learning network |
| CN108765291A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | Super resolution ratio reconstruction method based on dense neural network and two-parameter loss function |
| CN108765290A (en) * | 2018-05-29 | 2018-11-06 | 天津大学 | A kind of super resolution ratio reconstruction method based on improved dense convolutional neural networks |
| CN108765512A (en) * | 2018-05-30 | 2018-11-06 | 清华大学深圳研究生院 | A kind of confrontation image generating method based on multi-layer feature |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109903223A (en) | 2019-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109903223B (en) | Image super-resolution method based on dense connection network and generation type countermeasure network | |
| CN108537733B (en) | Super-resolution reconstruction method based on multi-path deep convolutional neural network | |
| CN109671023B (en) | Face image super-resolution secondary reconstruction method | |
| CN106683067B (en) | Deep learning super-resolution reconstruction method based on residual sub-images | |
| CN111275637B (en) | Attention model-based non-uniform motion blurred image self-adaptive restoration method | |
| CN107123089B (en) | Remote sensing image super-resolution reconstruction method and system based on depth convolution network | |
| CN113362223B (en) | Image super-resolution reconstruction method based on attention mechanism and two-channel network | |
| CN112734646B (en) | Image super-resolution reconstruction method based on feature channel division | |
| CN110175986B (en) | Stereo image visual saliency detection method based on convolutional neural network | |
| CN110136062B (en) | Super-resolution reconstruction method combining semantic segmentation | |
| CN111681166B (en) | Image super-resolution reconstruction method of stacked attention mechanism coding and decoding unit | |
| CN111242238B (en) | RGB-D image saliency target acquisition method | |
| CN109785236B (en) | Image super-resolution method based on super-pixel and convolutional neural network | |
| CN111754438B (en) | Underwater image restoration model based on multi-branch gating fusion and restoration method thereof | |
| CN112396607A (en) | Streetscape image semantic segmentation method for deformable convolution fusion enhancement | |
| CN112288632B (en) | Single image super-resolution method and system based on simplified ESRGAN | |
| CN110070574A (en) | A kind of binocular vision Stereo Matching Algorithm based on improvement PSMNet | |
| CN112837224A (en) | Super-resolution image reconstruction method based on convolutional neural network | |
| CN111833261A (en) | Image super-resolution restoration method for generating countermeasure network based on attention | |
| CN113139551A (en) | Improved semantic segmentation method based on deep Labv3+ | |
| CN115526777A (en) | Blind over-separation network establishing method, blind over-separation method and storage medium | |
| CN115713462A (en) | Super-resolution model training method, image recognition method, device and equipment | |
| CN114943646A (en) | Gradient weight loss and attention mechanism super-resolution method based on texture guidance | |
| CN110766609B (en) | Depth-of-field map super-resolution reconstruction method for ToF camera | |
| CN116485654A (en) | Lightweight single-image super-resolution reconstruction method combining convolutional neural network and transducer |
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 |