CN109191392A - A kind of image super-resolution reconstructing method of semantic segmentation driving - Google Patents
A kind of image super-resolution reconstructing method of semantic segmentation driving Download PDFInfo
- Publication number
- CN109191392A CN109191392A CN201810901713.3A CN201810901713A CN109191392A CN 109191392 A CN109191392 A CN 109191392A CN 201810901713 A CN201810901713 A CN 201810901713A CN 109191392 A CN109191392 A CN 109191392A
- Authority
- CN
- China
- Prior art keywords
- network
- resolution
- super
- semantic segmentation
- image
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
-
- 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
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to digital image processing techniques field, specially a kind of image super-resolution reconstructing method of semantic segmentation driving.The method of the present invention specifically includes: independently training image super-resolution network and semantic segmentation network model;Cascade the super-resolution network and semantic segmentation network of stand-alone training;Under the driving of semantic segmentation task, training super-resolution network;After the network processes that low-resolution image passes through task-driven, accurate semantic segmentation result is obtained.The experimental results showed that the present invention enables to super-resolution network to better adapt to segmentation task, clear, high resolution input picture is provided for semantic segmentation network, effectively improves the accuracy of separation of low-resolution image.
Description
Technical field
The invention belongs to digital image processing techniques fields, and in particular to a kind of image super-resolution reconstructing method, more
It says to body, is related to a kind of image super-resolution reconstructing method of semantic segmentation driving.
Background technique
Semantic segmentation is one of background task of computer vision field, and pixel is divided into difference according to different semantemes by it
Classification, have a wide range of applications in terms of automatic Pilot, image content.In recent years, depth convolutional neural networks
(deepconvolutional neural network, DCNN) not only has significant progress in image classification task, and
And in the task of some structurings output, such as semantic segmentation, it made breakthrough progress.
2015, Long et al.[1]It proposes FCN (fully convolutional neural network), for the first time will
DCNN is applied to the semantic segmentation task of Pixel-level classification.In order to keep receptive field, the pond layer in FCN is more, leads to spy
Sign figure resolution ratio is smaller, and segmentation result is coarse.Chen et al. proposes to improve characteristic pattern resolution ratio while not reducing receptive field
The method of Deeplab series[2-4], empty convolution is introduced, the output of network is optimized, in PASCAL VOC 2012[5]Survey
86.9% accuracy rate reached on examination collection.However, the wisp in segmented image is still very big chooses in semantic segmentation
War.
Image super-resolution reconstruct be it is a kind of effectively promote image resolution ratio, the technological means of rich image content, can be with
The visual effect of object effectively in enhancing wisp or low-resolution image.It is difficult simulation again based on early the reconstructing method of interpolation
Miscellaneous real scene.With the development of DCNN, also there are many ultra-resolution ratio reconstructing methods neural network based.
2015, Dong et al.[6]It is proposed SRCNN (Super-Resolution Convolutional Neural
Network), using low-resolution image as input, high-definition picture is as label, by optimization object function, allows DCNN
Learn the mapping relations between low-resolution image and high-definition picture.2016, Kim et al.[7]Deepen the network architecture,
It uses the image of interpolation as input, stacks multiple convolutional layers, and accelerate network convergence using the structure of residual error, achieve preferably
Quality reconstruction.
Above-mentioned ultra-resolution ratio reconstructing method is all but the figure seen of naked eyes for the purpose of the sensory effects for promoting human eye
The image that picture and machine are seen is not identical[8].Increase point of image rather than just the sensory effects of naked eyes for specific tasks
Resolution is beneficial to improve the effect of specific tasks.It is proposed a kind of ultra-resolution ratio reconstructing method of semantic segmentation driving for mentioning
The semantic segmentation accuracy of object has very strong practical value in high wisp or low-resolution image.
Summary of the invention
For overcome the deficiencies in the prior art, the purpose of the present invention is to provide a kind of image oversubscription of semantic segmentation driving
Resolution reconstructing method, allow super-resolution network can under the driving of semantic segmentation undated parameter, improve low-resolution image
The accuracy of semantic segmentation.
The image super-resolution reconstructing method of semantic segmentation driving provided by the invention, the specific steps are as follows:
(1) independently pre-training image super-resolution network and semantic segmentation network model
Use data setTraining super-resolution network, whereinIt is low-resolution image, as super-resolution
The input of rate network,It is high-definition picture, the label as training process;Super-resolution network used is an end
It can be VDSR to the network at end[7]、EDSR[9]Or SRCNN[6]Deng;
Use data setTraining semantic segmentation network, wherein IiFor the input of semantic segmentation network, MiFor pixel
Grade label, is image IiIn each pixel true classification;Semantic segmentation network used can be Deeplab[2-4]、FCN[1]
Or PSPnet[10]Deng;
(2) the super-resolution network and semantic segmentation network of stand-alone training are cascaded
Super-resolution network can be by low-resolution imageIt is mapped as high-definition pictureWherein θSR
For the parameter of super-resolution network;The output of super-resolution networkBy the input as semantic segmentation network, obtain
Obtain each pixel classifications result in super-resolution imageCascade network structure is constituted, wherein θseg
For the parameter of semantic segmentation network;
(3) under the driving of semantic segmentation task, training super-resolution network
Trim network parameter on the basis of pre-training model, with the loss function and semantic segmentation net of super-resolution network
The loss function of network instructs the update of the parameter of super-resolution network jointly, so that super-resolution network is for specific semantic point
The task of cutting is adjusted;
(4) low-resolution image obtains accurate semantic segmentation result after the network processes of task-driven
For the semantic segmentation task of low-resolution image, low-resolution image is first input to the semanteme point of training completion
It cuts in the super-resolution network model of driving, reconstructs high-definition picture, then the high-definition picture of reconstruct is input to semanteme
Divide in network, obtains accurate segmentation result.
Further, in step (1), the data set of training super-resolution networkAcquisition methods are as follows:
By high-definition pictureDown-sampling according to a certain percentage obtains low-resolution imageFor amateur use
In the higher image data set of the resolution ratio of super resolution task, the data of super resolution task can be constructed in this way
Collection.The data set of image super-resolution reconstruct is different from the data set of other tasks, such as object detection data set, image classification number
According to collection etc., do not need manually to mark, therefore super-resolution reconstruction can combine with other tasks, while not needing standard
For the data set of multiple tasks, make it possible the super-resolution reconstruction of task-driven.
Further, in step (1), the method for two kinds of network stand-alone trainings are as follows:
With two kinds of data set training super-resolution networks, network first is trained with the data set of common super resolution task,
After convergence, then with semantic partitioned data set trim network;
With the semantic segmentation data set training semantic segmentation network containing Pixel-level mark of standard.
The data set of common super resolution task can be DIV2K[11], 91 pictures[12]Deng;Semantic segmentation data set
It can be PASCAL VOC 2012[5]、PASCAL context[13]Or Cityscapes[14]Deng.
Further, in step (2), the parameter of cascade network by two stand-alone trainings model initialization, wherein semantic
The parameter of parted pattern part will be fixed, the loss for the semantic segmentation that the high-definition picture for calculating reconstruct generates, in advance
Trained semantic segmentation network provides correct guidance for the parameter update of super-resolution network, therefore one more accurate semantic
It is most important in cascade network to divide network model.
Further, in step (3), the loss function are as follows:
The loss function of super-resolution network are as follows:
Wherein, N is amount of images.
The loss function of semantic segmentation network are as follows:
Wherein,
L is the set of the classification of pixel,For the pixel for belonging to l class in i-th of image,For l class pixel
Quantity, u are the positions of pixel.
In order to make super-resolution network can adapt to semantic segmentation task, rather than it is provided solely for preferable visual quality,
The loss function of the loss function of super-resolution network and semantic segmentation network is combined, as final loss function, so
The objective function that parameter updates are as follows:
Wherein, α, β are used to balance the contribution of two kinds of loss functions, and the selection of α, β can adjust according to demand, ordinary circumstance
Under, α is smaller with respect to β, and the high-definition picture visual effect of reconstruct is poorer, but semantic segmentation accuracy is higher;α is bigger with respect to β,
It is then opposite;It is recommended that α, β ratio between two are taken as (0.5-1): 1, preferably 1:1.
Although the present invention uses the form of two cascades, minimizing gradient that loss function generates still can be with
It travels in super-resolution network, loss function is to θSRGradient are as follows:
The beneficial effects of the present invention are: the present invention makes for specific semantic segmentation task training super-resolution network
The purpose for obtaining super-resolution reconstruction is no longer only to provide the high-definition picture that visual quality is higher, details is richer, but
Export high-definition picture that is richer for semantic segmentation Web content, being more advantageous to extraction feature.In the present invention, semanteme point
The ultra-resolution ratio reconstructing method frame for cutting driving is simple, is easily achieved, and can be used as a kind of pre-treatment and is widely used, improves low
The semantic segmentation accuracy of image in different resolution.
Detailed description of the invention
Fig. 1 is network frame figure of the invention.
Fig. 2 is that the segmentation effect of the high-definition picture reconstructed using method of the invention and other methods compares (4 times of weights
Structure).
Specific embodiment
Embodiment of the present invention is described in detail below, but protection scope of the present invention is not limited to the implementation
Example.
Using VDSR as super-resolution network, Deeplab-V2 does 4 times and 8 times reconstruct as semantic segmentation network respectively,
Low-resolution image is obtained by high-definition picture down-sampling, the specific steps are as follows:
(1) stand-alone training super-resolution network VDSR and semantic segmentation network Deeplab-V2.With DIV2K and PASCAL
VOC 2012 trains super-resolution network;With the training semantic segmentation network of PASCAL VOC 2012;
(2) the super-resolution network and semantic segmentation network for cascading stand-alone training, with the parameter initialization grade in step (1)
The parameter of corresponding part in networking network;
(3) under the driving of semantic segmentation task, training super-resolution network, the weight of loss function α, β be set as 1:1 or
0.5:1;
(4) after the network processes that low-resolution image passes through task-driven, accurate semantic segmentation result is obtained.
The image for image and the other methods reconstruct that the present invention reconstructs, after semantic segmentation network processes, the accuracy of separation
Compare as shown in table 1.As can be seen that under different reconstruct multiples, the essence of the high-definition picture segmentation of method reconstruct of the invention
Exactness is significantly larger than other methods.
In addition, give α: 0.5 in Fig. 2, when β: 1, in 4 times of reconstruct, the method for the present invention and other methods reconstructed image point
The intuitively comparing of effect is cut, wherein Fig. 2 (a) is high-definition picture and semantic segmentation label;Fig. 2 (b) is bicubic difference weight
The image and semantic segmentation result of structure;Fig. 2 (c) is the reconstructed image and semantic segmentation knot of the super-resolution network of stand-alone training
Fruit;Fig. 2 (d) is the reconstructed image and segmentation result of the super-resolution network of semantic segmentation driving.As can be seen that side of the invention
The segmentation result that method generates is most accurate.
The comparison of the accuracy of separation of 1 distinct methods reconstructed image of table
Bibliography
[1]J.Long,E.Shelhamer and T.Darrell,“Fully convolutional networks for
semantic segmentation,”IEEE Conference on Computer Vision and Pattern
Recognition(CVPR),pp.3431-3440,2015.(FCN)
[2]L.Chen,G.Papandreou,and et al.,“Semantic image segmentation with
deep convolutional nets and fully connected CRFs,”International Conference on
Learning Representations(ICLR),2015.(DeepLab-V1)
[3]L.Chen,G.Papandreou,and et al.,“DeepLab:Semantic image
segmentation with deep convolutional nets,atrous convolution,and fully
connected CRFs,”IEEE Transactions on Pattern Analysis and Machine
Intelligence(TPMAI),vol.40,pp.834-848,2018.(DeepLab-V2)
[4]L.Chen,G.Papandreou,and et al.“Rethinking atrous convolution for
semantic image segmentation,”arXiv:1706.05587,2017.(DeepLab-V3)
[5]M.Everingham,S.Eslami,and et al.,“The pascal visual object classes
challenge:a retrospective,”International Journal of Computer Vision(IJCV),
vol.111,no.1,pp.98-136,2014.(PASCAL VOC 2012)
[6]C.Dong,C.C.Loy,K.He,and X.Tang.“Image super-resolution using deep
convolutional networks,”IEEE Transactions on Pattern Analysis and Machine
Intelligence(TPAMI),vol.38,no.2,pp.295-307,2015.(SRCNN)
[7]J.Kim,J.Lee,and et al.“Accurate image super-resolution using very
deep convolutional networks,”IEEE Conference on Computer Vision and Pattern
Recognition(CVPR),pp.1646-1654,2016.(VDSR)
[8]C.Xie,J.Wang,Z.Zhang,Y.Zhou,L.Xie and A.Yuille,“Adversarial
Examples for Semantic Segmentation and Object Detection,”IEEE International
Conference on Computer Vision(ICCV),pp.1378-1387,2017.
[9]B.Lim,S.Son,H.Kim,S.Nah and K.M.Lee,“Enhanced Deep Residual
Networks for Single Image Super-Resolution,”IEEE Conference on Computer
Vision and Pattern Recognition Workshops(CVPRW),pp.1132-1140,2017.(EDSR)
[10]H.Zhao,J.Shi,X.Qi,X.Wang,and J.Jia,“Pyramid Scene Parsing
Network,”IEEE Conference on Computer Vision and Pattern Recognition(CVPR),
pp.2881-2890,2017.(PSPnet)
[11]R.Timofte,E.Agustsson,L.Van Gool,M.-H.Yang,L.Zhang,et al.,“Ntire
2017 challenge on single image superresolution:Methods and results,”IEEE
Conference on Computer Vision and Pattern Recognition Workshops(CVPRW),2017.
(DIV2K)
[12]J.Yang,J.Wright,T.S.Huang,and Y.Ma,“Image super-resolution via
sparse representation,”IEEE Transactions on Image Processing,pp.2861-2873,
2010.(91 images)
[13]R.Mottaghi,X.Chen,X.Liu,and et al.,“The role of context for
object detection and semantic segmentation in the wild,”IEEE Conference on
Computer Vision and Pattern Recognition(CVPR),2014.(PASCAL context)
[14]M.Cordts,M.Omran,S.Ramos,and et al.,“The cityscapes dataset for
semantic urban scene understanding,”IEEE Conference on Computer Vision and
Pattern Recognition(CVPR),2016.(Cityscapes)。
Claims (7)
1. a kind of image super-resolution reconstructing method of semantic segmentation driving, which is characterized in that specific step is as follows:
(1) independently pre-training image super-resolution network and semantic segmentation network model
Use data setTraining super-resolution network, whereinIt is low-resolution image, as super-resolution net
The input of network,It is high-definition picture, the label as training process;
Use data setTraining semantic segmentation network, wherein IiFor the input of semantic segmentation network, MiFor Pixel-level
Label indicates image IiIn each pixel true classification;
(2) the super-resolution network and semantic segmentation network of stand-alone training are cascaded
Super-resolution network is by low-resolution imageIt is mapped as high-definition pictureWherein θSRFor oversubscription
The parameter of resolution network;The output of super-resolution networkAs the input of semantic segmentation network, surpassed
Each pixel classifications result in image in different resolutionCascade network structure is constituted, wherein θseg
For the parameter of semantic segmentation network;
(3) under the driving of semantic segmentation task, training super-resolution network
Trim network parameter on the basis of pre-training model, with the loss function of super-resolution network and semantic segmentation network
Loss function instructs the update of the parameter of super-resolution network jointly, so that super-resolution network is appointed for specific semantic segmentation
Business is adjusted;
(4) network processes of the low-resolution image Jing Guo task-driven obtain accurate semantic segmentation result
For the semantic segmentation task of low-resolution image, first the semantic segmentation that low-resolution image is input to training completion is driven
In dynamic super-resolution network model, high-definition picture is reconstructed, then the high-definition picture of reconstruct is input to semantic segmentation
In network, accurate segmentation result is obtained.
2. the method according to claim 1, wherein training the data set of super-resolution network in step (1)Acquisition methods are as follows:
By high-definition pictureDown-sampling according to a certain percentage obtains low-resolution imageOversubscription is used for for amateur
The higher image data set of the resolution ratio of resolution task, all in accordance with the data set of the method building super resolution task.
3. according to the method described in claim 2, it is characterized in that, in step (1), the method for two kinds of network stand-alone trainings are as follows:
It is restrained with two kinds of data set training super-resolution networks first with the data set training network of common super resolution task
Afterwards, then with semantic partitioned data set trim network;
With the semantic segmentation data set training semantic segmentation network containing Pixel-level mark of standard.
4. requiring the method according to claim 3, which is characterized in that the data set of the common super resolution task
For DIV2K or 91 pictures;The semantic segmentation data set be PASCAL VOC 2012, PASCAL context or
Cityscapes。
5. the method according to claim 1, wherein the parameter of cascade network is by two independent instructions in step (2)
Experienced model parameter initialization;Wherein the parameter of semantic segmentation model part is fixed, for calculating the high resolution graphics of reconstruct
As the loss of the semantic segmentation generated, the semantic segmentation network of pre-training provides correctly for the parameter update of super-resolution network
Guidance.
6. the method according to claim 1, wherein in step (3), the loss function are as follows:
The loss function of super-resolution network are as follows:
Wherein, N is amount of images;
The loss function of semantic segmentation network are as follows:
Wherein,
L is the set of the classification of pixel,For the pixel for belonging to l class in i-th of image,For the quantity of l class pixel,
U is the position of pixel;
The loss function of the loss function of super-resolution network and semantic segmentation network is combined, as final loss function,
So the objective function that parameter updates are as follows:
Wherein, α, β are used to balance the contribution of two kinds of loss functions;α, β ratio between two are taken as (0.5-1): 1.
7. the method according to claim 1, wherein the super-resolution network is a net end to end
Network is VDSR, EDSR or SRCNN;Semantic segmentation network is Deeplab, FCN or PSPnet.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810901713.3A CN109191392B (en) | 2018-08-09 | 2018-08-09 | Image super-resolution reconstruction method driven by semantic segmentation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810901713.3A CN109191392B (en) | 2018-08-09 | 2018-08-09 | Image super-resolution reconstruction method driven by semantic segmentation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109191392A true CN109191392A (en) | 2019-01-11 |
CN109191392B CN109191392B (en) | 2021-06-04 |
Family
ID=64921175
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810901713.3A Active CN109191392B (en) | 2018-08-09 | 2018-08-09 | Image super-resolution reconstruction method driven by semantic segmentation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109191392B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009565A (en) * | 2019-04-04 | 2019-07-12 | 武汉大学 | A kind of super-resolution image reconstruction method based on lightweight network |
CN110136062A (en) * | 2019-05-10 | 2019-08-16 | 武汉大学 | A kind of super resolution ratio reconstruction method of combination semantic segmentation |
CN110837811A (en) * | 2019-11-12 | 2020-02-25 | 腾讯科技(深圳)有限公司 | Method, device and equipment for generating semantic segmentation network structure and storage medium |
CN112419158A (en) * | 2020-12-07 | 2021-02-26 | 上海互联网软件集团有限公司 | Image video super-resolution and super-definition reconstruction system and method |
US20210209732A1 (en) * | 2020-06-17 | 2021-07-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Face super-resolution realization method and apparatus, electronic device and storage medium |
CN113538227A (en) * | 2020-04-20 | 2021-10-22 | 华为技术有限公司 | Image processing method based on semantic segmentation and related equipment |
CN113657388A (en) * | 2021-07-09 | 2021-11-16 | 北京科技大学 | Image semantic segmentation method fusing image super-resolution reconstruction |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107481188A (en) * | 2017-06-23 | 2017-12-15 | 珠海经济特区远宏科技有限公司 | A kind of image super-resolution reconstructing method |
WO2018067258A1 (en) * | 2016-10-06 | 2018-04-12 | Qualcomm Incorporated | Neural network for image processing |
CN108140130A (en) * | 2015-11-05 | 2018-06-08 | 谷歌有限责任公司 | The bilateral image procossing that edge perceives |
CN108259994A (en) * | 2018-01-15 | 2018-07-06 | 复旦大学 | A kind of method for improving video spatial resolution |
-
2018
- 2018-08-09 CN CN201810901713.3A patent/CN109191392B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108140130A (en) * | 2015-11-05 | 2018-06-08 | 谷歌有限责任公司 | The bilateral image procossing that edge perceives |
WO2018067258A1 (en) * | 2016-10-06 | 2018-04-12 | Qualcomm Incorporated | Neural network for image processing |
CN107481188A (en) * | 2017-06-23 | 2017-12-15 | 珠海经济特区远宏科技有限公司 | A kind of image super-resolution reconstructing method |
CN108259994A (en) * | 2018-01-15 | 2018-07-06 | 复旦大学 | A kind of method for improving video spatial resolution |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110009565A (en) * | 2019-04-04 | 2019-07-12 | 武汉大学 | A kind of super-resolution image reconstruction method based on lightweight network |
CN110136062A (en) * | 2019-05-10 | 2019-08-16 | 武汉大学 | A kind of super resolution ratio reconstruction method of combination semantic segmentation |
CN110136062B (en) * | 2019-05-10 | 2020-11-03 | 武汉大学 | Super-resolution reconstruction method combining semantic segmentation |
CN110837811A (en) * | 2019-11-12 | 2020-02-25 | 腾讯科技(深圳)有限公司 | Method, device and equipment for generating semantic segmentation network structure and storage medium |
CN110837811B (en) * | 2019-11-12 | 2021-01-05 | 腾讯科技(深圳)有限公司 | Method, device and equipment for generating semantic segmentation network structure and storage medium |
CN113538227A (en) * | 2020-04-20 | 2021-10-22 | 华为技术有限公司 | Image processing method based on semantic segmentation and related equipment |
CN113538227B (en) * | 2020-04-20 | 2024-04-12 | 华为技术有限公司 | Image processing method based on semantic segmentation and related equipment |
US20210209732A1 (en) * | 2020-06-17 | 2021-07-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Face super-resolution realization method and apparatus, electronic device and storage medium |
US11710215B2 (en) * | 2020-06-17 | 2023-07-25 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Face super-resolution realization method and apparatus, electronic device and storage medium |
CN112419158A (en) * | 2020-12-07 | 2021-02-26 | 上海互联网软件集团有限公司 | Image video super-resolution and super-definition reconstruction system and method |
CN113657388A (en) * | 2021-07-09 | 2021-11-16 | 北京科技大学 | Image semantic segmentation method fusing image super-resolution reconstruction |
CN113657388B (en) * | 2021-07-09 | 2023-10-31 | 北京科技大学 | Image semantic segmentation method for super-resolution reconstruction of fused image |
Also Published As
Publication number | Publication date |
---|---|
CN109191392B (en) | 2021-06-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109191392A (en) | A kind of image super-resolution reconstructing method of semantic segmentation driving | |
CN109389556A (en) | The multiple dimensioned empty convolutional neural networks ultra-resolution ratio reconstructing method of one kind and device | |
CN113362223B (en) | Image super-resolution reconstruction method based on attention mechanism and two-channel network | |
CN108492248A (en) | Depth map super-resolution method based on deep learning | |
CN107844795B (en) | Convolutional neural network feature extraction method based on principal component analysis | |
CN110634108B (en) | Composite degraded network live broadcast video enhancement method based on element-cycle consistency confrontation network | |
CN109064396A (en) | A kind of single image super resolution ratio reconstruction method based on depth ingredient learning network | |
CN109035149A (en) | A kind of license plate image based on deep learning goes motion blur method | |
CN108259994B (en) | Method for improving video spatial resolution | |
CN107563965A (en) | Jpeg compressed image super resolution ratio reconstruction method based on convolutional neural networks | |
CN107993238A (en) | A kind of head-and-shoulder area image partition method and device based on attention model | |
CN110163801A (en) | A kind of Image Super-resolution and color method, system and electronic equipment | |
Chen et al. | Single image super-resolution using deep CNN with dense skip connections and inception-resnet | |
CN109920012A (en) | Image colorant system and method based on convolutional neural networks | |
CN107784628A (en) | A kind of super-resolution implementation method based on reconstruction optimization and deep neural network | |
CN116091886A (en) | Semi-supervised target detection method and system based on teacher student model and strong and weak branches | |
CN112288630A (en) | Super-resolution image reconstruction method and system based on improved wide-depth neural network | |
CN111127331A (en) | Image denoising method based on pixel-level global noise estimation coding and decoding network | |
CN110751271B (en) | Image traceability feature characterization method based on deep neural network | |
CN112580473A (en) | Motion feature fused video super-resolution reconstruction method | |
CN112364838A (en) | Method for improving handwriting OCR performance by utilizing synthesized online text image | |
CN115861614A (en) | Method and device for automatically generating semantic segmentation graph based on down jacket image | |
Gao et al. | Learning to Incorporate Texture Saliency Adaptive Attention to Image Cartoonization. | |
CN111105354A (en) | Depth image super-resolution method and device based on multi-source depth residual error network | |
CN109993701A (en) | A method of the depth map super-resolution rebuilding based on pyramid structure |
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 |