CN110351568A - A kind of filtering video loop device based on depth convolutional network - Google Patents
A kind of filtering video loop device based on depth convolutional network Download PDFInfo
- Publication number
- CN110351568A CN110351568A CN201910511992.7A CN201910511992A CN110351568A CN 110351568 A CN110351568 A CN 110351568A CN 201910511992 A CN201910511992 A CN 201910511992A CN 110351568 A CN110351568 A CN 110351568A
- Authority
- CN
- China
- Prior art keywords
- video
- filtering
- network
- image
- frame
- 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
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/50—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding
- H04N19/503—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding involving temporal prediction
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/50—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding
- H04N19/593—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding involving spatial prediction techniques
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/70—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals characterised by syntax aspects related to video coding, e.g. related to compression standards
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/80—Details of filtering operations specially adapted for video compression, e.g. for pixel interpolation
- H04N19/82—Details of filtering operations specially adapted for video compression, e.g. for pixel interpolation involving filtering within a prediction loop
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The invention discloses a kind of method for filtering video loop based on depth convolutional network, step 1, production are used for the training dataset of loop filtering network training;Step 2, building are used for the network model of video filtering;Two parts model for the filtering of intra prediction frame and MB of prediction frame filtering is respectively trained as the training set of network in step 3, the training dataset for obtaining step 1, this two parts model constitutes video filtering network model;To minimize loss function as optimization aim, video filtering network model is trained;Video filtering network model trained in step 3 is integrated into Video coding software by step 4, to complete entire video coding process, obtains the reconstructed frame after video filtering network.The present invention improves the picture quality of video reconstruction frames compared with traditional filtering method, improves the accuracy of inter-prediction, improves code efficiency, and filtered picture frame has higher reconstruction quality, greatly improves video coding efficiency.
Description
Technical field
The invention belongs to computer vision fields and field of video encoding, and in particular to a kind of based on depth convolutional network
Filtering video loop device.
Background technique
As state-of-the-art video encoding standard, HEVC (High Efficiency Video Coding) relatively before view
Frequency coding standard compression efficiency is promoted huge.Although the loop filter of HEVC has been obviously improved reconstructed frame in Video coding
Quality, but there is also many limitations for the loop filtering technology of HEVC.Caused for example, compressed video still exists by quantization
Unpleasant image fault.With the development of internet and new media, more and more views all can be played and propagated daily
Frequently.Therefore, higher there is an urgent need to provide quality, the smaller video of volume.
All existing coding standards all use hybrid encoding frame, including predict, change and quantization waited within the frame/frames
Journey.Motion compensation when due to the coarse quantization of image frequency domain coefficient and inter-prediction, it is inevitable in video reconstruction frames
Cause a large amount of distortions such as blocking artifact, ringing effect, image be fuzzy.Loop filter has become ten in Video coding
Divide important technology, the quality of decoding end decoding frame not only can be directly improved by reducing compression artefacts, can also be improved
The quality of inter prediction reference frame reduces number of coded bits by motion compensation.
Deblocking filtering is intended to reduce the blocking artifact of reconstructed frame, and when quantization parameter is smaller, apparent block is not present in image
When effect, deblocking filtering is difficult to bring effective increased quality.Pixel adaptive equalization carries out the pixel in reconstructed frame thick
Granularity processing, bring reconstructed frame increased quality is limited, and needs to transmit a large amount of adaptive equalization parameters to decoding end, increases
Number of coded bits.This is difficult to meet the needs of people are for low bit, high-quality video.
Summary of the invention
The present invention is directed to be directed to the limitation of existing filtering video loop device, a kind of view based on depth convolutional network is proposed
Frequency loop circuit filtering method realizes distorted image and original image using the network model obtained based on the study of depth convolutional network
Between more accurate Nonlinear Mapping, and take square filter and adaptive pixel compensation instead of in conventional filter,
To realize video filtering.
A kind of method for filtering video loop based on depth convolutional network of the invention, comprising the following steps:
Step 1, selection image data set and sets of video data concentrate unpressed original image and video to carry out data
Compression, the distorted image and original image that obtain after image data set is compressed are fabricated to the training of intra prediction frame filter network
Data set;By after video compress obtained distorted image frame and original image frame be fabricated to the training of inter-prediction filter network
Data set;
Step 2, building are used for the network model of video filtering, which includes multiple convolutional layers, active coating and normalization
Layer, there is across convolutional layer short circuit between different convolutional layers and connects;First layer convolutional layer characterizes input picture at characteristic pattern, finally
One layer of convolutional layer then rebuilds the residual error between compression image and original image, and middle layer is the non-thread of distorted image and original image
Property mapping residual error unit;
Step 3, two kinds of training datasets for obtaining step 1 are respectively trained pre- for the filtering of intra prediction frame and interframe
The network model of two kinds of different characteristics of frame filtering is surveyed, both models are respectively filtered frame different types of in video;
To minimize loss function as optimization aim, video filtering network model is trained, the loss function of two kinds of models is identical,
Only trained data set is different, the two model characteristics differences trained, filters respectively for different types of video frame.It gives
Determine training setN is the number of training concentration training luminance block, xiFor the luminance block being distorted after compression, yiIt is uncompressed
Original brightness block;Then the loss function mathematical expression of brightness network is as follows:
Wherein, Θ indicates that the parameter set of network, the loss function pass through backpropagation small lot stochastic gradient descent algorithm
Optimization, the initial weight of network model use the MSRA initial method suitable for ReLU activation primitive;
Video filtering network model trained in step 3 is integrated into Video coding software by step 4, removes coding
Deblocking filtering and adaptive pixel compensation module in software are obtained with completing entire video coding process through video filtering
Reconstructed frame after network.
The present invention improves the picture quality of video reconstruction frames compared with traditional filtering method, the high quality after network filtering
Video frame can be further used as reference picture, improve the accuracy of inter-prediction;
In addition, the present invention does not need to encode and transmit any parameter to decoding end, bit needed for reducing encoded video
Number;Filtered picture frame has higher reconstruction quality, and does not need that any parameter is written into coded bit stream, greatly
Ground improves video coding efficiency.
Detailed description of the invention
Fig. 1 is a kind of method for filtering video loop overall flow figure based on depth convolutional network of the invention;
Fig. 2 is the structural schematic diagram of the embodiment of the present invention network model.
Specific embodiment
Technical solution of the present invention is described in detail with reference to the accompanying drawings and examples.
Step 1, for the intra prediction frame distorted characteristic different with MB of prediction frame, the present embodiment has chosen two differences
Data set make the training set for intra prediction frame and inter-prediction frame network respectively.
For intra prediction frame, the present embodiment selection includes the UCID (Uncompressed of 1338 natural images
Colour Image Database) image data set, every image uses HEVC reference software (All in full frame in data set
Intra), close under the configuration of deblocking filtering and adaptive pixel compensation and compressed, compressed image is divided into 35x35
Block of pixels, and be fabricated to together with original image the training set of intra prediction frame filter network.
For inter-prediction, the present embodiment has selected the HEVC standard video of 15 different resolutions, each video every five
Frame chooses a frame and extracts 50 frames altogether, to this 15 video sequences using HEVC reference software in random access (Random
Access), close under the configuration of deblocking filtering and adaptive pixel compensation and compressed, compressed image is divided into 35 ×
35 block of pixels, and it is fabricated to together with original image the training set of MB of prediction frame filter network.
Step 2, in view of the balance between network depth and computation complexity, the network of the present embodiment includes 5 residual errors
Unit, continues growing network depth and is promoted less to network performance, but will increase network query function complexity by totally 12 layers.Network
First layer convolutional layer characterizes input picture at characteristic pattern, and the number of characteristic pattern is related with the number of this layer of convolution kernel.Last
The convolutional layer of layer then rebuilds the residual error between compression image and original image, and network middle section is 5 residual units, study distortion
The Nonlinear Mapping of image and original image.Specifically, two kinds of residual error learning strategies of Web vector graphic (i.e. external residual error study
Strategy and internal residual error study), to overcome the problems, such as that it is trained that deep layer network is difficult to.A kind of residual error learning strategy prediction input picture
Residual error between (distorted image) and output image (original image), rather than directly prediction output image, the residual error learn plan
Slightly it is referred to as external residual error study.It is as follows to define residual image expression formula:
R=Y-X
Wherein, X indicates compressed reconstruction image, and Y indicates unpressed original image.It is expected that prediction residual image R and
It is not original image Y, therefore loss function becomesHere F (x) indicates network output valve.Another residual error
Practise strategy be then between network internal adjacent layer by short circuit connection realization identical mapping, the network every two convolutional layer it
Between introduce a short circuit connection and realize identical mapping, a structure the smallest in this way is referred to as residual unit, optimization residual error mapping
It is more easier than original mappings, which is referred to as internal residual error study.Except the last layer is for residual in this method
Outside the convolutional layer that difference image is rebuild, each layer of convolutional layer all using using preceding to activation structure, the structure by normalized function and
Activation primitive is placed on before weight layer, and this forward direction activation primitive is more easier network training and improves the extensive energy of network
Power.It is then as follows using the preceding residual unit mathematic(al) representation to activation:
Dl+1=R (Dl)=h (Dl)+F(Dl,Wl)
Wherein, DlIndicate the input feature vector of first of residual unit, Wl={ Wl,k|1≤k≤KIndicate and first of residual unit phase
The weight (including biasing) of pass, K indicate the number of plies in residual unit, and F indicates residual error function, and h indicates identical mapping function: h
(Dl)=Dl。
Internal residual error learning strategy effectively overcomes the problem of deep layer network is degenerated, and external residual error learning strategy is greatly
It accelerates the convergence rate of network and effectively raises the quality of output image.In addition, this method is whole by trained model
It closes in coding and decoding video software, does not need any additional pretreatment and last handling process, realize video volume end to end
Decoding, facilitates the deployment and use of coding/decoding system.
In the construction process of network model, it is contemplated that the volume and computation complexity of network model, it is necessary to as far as possible
Model parameter amount is reduced, all convolutional layers all use 3 × 3 convolution kernel thus, in addition to the last layer convolutional layer, all volumes
The convolution kernel number of lamination is 64.The beginning of network model introduces one layer of convolutional layer, which characterizes input picture at feature
Figure, the number of characteristic pattern are related with the number of this layer of convolution kernel.Then several residual units are stacked.The last layer volume
Lamination rebuilds the residual error between compression image and original image.Assuming that x is the input of network, then D0=faIt (x) is first layer convolutional layer
Output, faFor the function of first layer convolutional layer.The output of m-th of residual unit are as follows:
Dm=R(m)(fa(x))=R (R (... R (fa(x))...))
When the residual unit number in network is M, which can indicate as follows:
Y=N (x)=fl(DM)+x=fl(R(R(...R(fa(x))...)))+x.
Wherein, fiFor the function of the last layer convolutional layer.
In view of the degenerate problem of deep layer network, short-circuit connection is added in this example between every two convolutional layer, is had
The deep layer network that overcomes of effect is difficult to trained problem, solves network degenerate problem, enhances the learning ability of network.Network
Residual error after study compression between distorted image and original image, then further speed up the convergence rate of network, improves reconstruction image
Quality.The block of pixels (64 × 64) that the input picture of network is divided into fixed size is handled respectively, on the one hand can be reduced
On the other hand physical memory needed for network query function can guarantee that the video of different resolution is all filtered using consolidated network
Wave.The activation primitive of each convolutional layer, which is selected, in this example corrects linear unit R eLU function:
F (x)=max (0, x)
ReLU activation primitive can preferably transmit the gradient of loss function in depth convolutional network, help model faster
Convergence, be effectively prevented gradient disperse.
Two different network models are respectively trained with two training sets made in step 1 in step 3, are respectively used to
Intra prediction frame and MB of prediction frame filtering.Training platform uses Caffe, criticizes and is dimensioned to 64, activation primitive ReLU, just
Beginning method uses MSRA.Loss function is
Momentum parameter is 0.9, and weight decays to 0.0001, and optimization method uses stochastic gradient descent, and initial learning rate is set
It is set to 0.1, the every training 10 of training set takes turns learning rate and is kept to original 0.25, trains 80 wheels altogether, and adjustable gradient is cut out for accelerating
The convergence rate of network.Design parameter in this example when network training is summarized as follows:
Parameter name | Parameter value |
Input picture block size (block size) | 35×35 |
It criticizes size (batch size) | 64 |
The number of iterations (epochs) | 80epochs |
Momentum parameter (momentum parameter) | 0.9 |
Weight decays (weight decay) | 0.0001 |
Learning rate (learning rate) | 0.1, every 10 epochs are multiplied by 0.25 |
Trained network model is integrated into HEVC reference software by step 4, first the side of going in removing encoding software
Block filtering and adaptive pixel compensation module are called by the image segmentation obtained by encoding software at 64 × 64 block of pixels
The C++ interface of Caffe judges that filtering image is intra prediction frame or MB of prediction frame, according to different frame type uses pair
Trained network model is answered to be filtered unfiltered 64 × 64 block of pixels, finally by all filtered block of pixels groups
At reconstruction image.The process realizes encoding software encoding and decoding end to end processing, does not need any pretreatment and last handling process,
It is easy to the deployment and actual use of coding/decoding system.
Network model in this example uses deep layer network, while using two different residual error learning strategies, and is directed to frame
It is different for being respectively trained two to make two different training sets for the interior prediction frame distorted characteristic different with MB of prediction frame
Network model realizes accurately Nonlinear Mapping between distortion pixel and original pixels, for different resolution, different content
Video all has good versatility.
Claims (3)
1. a kind of method for filtering video loop based on depth convolutional network, which is characterized in that this method includes below scheme:
Step 1, selection image data set and sets of video data concentrate unpressed original image and video to press data
Contracting, the distorted image and original image obtained after image data set is compressed are fabricated to the training number of intra prediction frame filter network
According to collection;By after video compress obtained distorted image frame and original image frame be fabricated to the training number of inter-prediction filter network
According to collection;
Step 2, building are used for the network model of video filtering, which includes multiple convolutional layers, active coating and normalize layer, no
It is connected with there is across convolutional layer short circuit between convolutional layer;First layer convolutional layer characterizes input picture at characteristic pattern, the last layer
Convolutional layer then rebuild the residual error between compression image and original image, middle layer is that distorted image and the non-linear of original image reflect
The residual error unit penetrated;
Step 3, two kinds of training datasets for obtaining step 1 are respectively trained for the filtering of intra prediction frame and MB of prediction frame
The network model of two kinds of different characteristics of filtering, both models are respectively filtered frame different types of in video;With most
Smallization loss function is optimization aim, is trained to video filtering network model, the loss function of two kinds of models is identical, only
Trained data set is different, the two model characteristics differences trained, filters respectively for different types of video frame.Given instruction
Practice collectionN is the number of training concentration training luminance block, xiFor the luminance block being distorted after compression, yiFor unpressed original
Beginning luminance block;Then the loss function mathematical expression of brightness network is as follows:
Wherein, Θ indicates that the parameter set of network, the loss function are optimized by backpropagation small lot stochastic gradient descent algorithm,
The initial weight of network model uses the MSRA initial method suitable for ReLU activation primitive;
Video filtering network model trained in step 3 is integrated into Video coding software by step 4, removes encoding software
In deblocking filtering and adaptive pixel compensation module obtained with completing entire video coding process through video filtering network
Reconstructed frame afterwards.
2. a kind of method for filtering video loop based on depth convolutional network as described in claim 1, which is characterized in that described
For residual unit in step 2 using two kinds of external residual error learning strategies, the expression formula for defining residual image R is as follows:
R=Y-X
Wherein, X indicates compressed reconstruction image, and Y indicates unpressed original image.
3. a kind of method for filtering video loop based on depth convolutional network as described in claim 1, which is characterized in that described
For residual unit in step 2 using two kinds of external residual error learning strategies, the expression formula for defining residual image R is as follows:
Dl+1=R (Dl)=h (Dl)+F(Dl,Wl)
Wherein, DlIndicate the input feature vector of first of residual unit, Wl={ Wl,k|1≤k≤KIndicate relevant to first of residual unit
Weight (including biasing), K indicate the number of plies in residual unit, and F indicates residual error function, and h indicates identical mapping function: h (Dl)=
Dl。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910511992.7A CN110351568A (en) | 2019-06-13 | 2019-06-13 | A kind of filtering video loop device based on depth convolutional network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910511992.7A CN110351568A (en) | 2019-06-13 | 2019-06-13 | A kind of filtering video loop device based on depth convolutional network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110351568A true CN110351568A (en) | 2019-10-18 |
Family
ID=68181931
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910511992.7A Pending CN110351568A (en) | 2019-06-13 | 2019-06-13 | A kind of filtering video loop device based on depth convolutional network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110351568A (en) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110933422A (en) * | 2019-11-22 | 2020-03-27 | 南京信息工程大学 | HEVC loop filtering method based on EDCNN |
CN111405283A (en) * | 2020-02-20 | 2020-07-10 | 北京大学 | End-to-end video compression method, system and storage medium based on deep learning |
CN111541894A (en) * | 2020-04-21 | 2020-08-14 | 电子科技大学 | Loop filtering method based on edge enhancement residual error network |
CN111866521A (en) * | 2020-07-09 | 2020-10-30 | 浙江工商大学 | Video image compression artifact removing method combining motion compensation and generation type countermeasure network |
CN113012073A (en) * | 2021-04-01 | 2021-06-22 | 清华大学 | Training method and device for video quality improvement model |
CN113132729A (en) * | 2020-01-15 | 2021-07-16 | 北京大学 | Loop filtering method based on multiple reference frames and electronic device |
CN113259671A (en) * | 2020-02-10 | 2021-08-13 | 腾讯科技(深圳)有限公司 | Loop filtering method, device and equipment in video coding and decoding and storage medium |
CN113452998A (en) * | 2020-03-25 | 2021-09-28 | 杭州海康威视数字技术股份有限公司 | Decoding, encoding and decoding method, device and equipment thereof |
CN113613000A (en) * | 2021-08-20 | 2021-11-05 | 天津大学 | Intelligent multi-resolution depth video intra-frame prediction method |
CN113852850A (en) * | 2020-11-24 | 2021-12-28 | 广东朝歌智慧互联科技有限公司 | Audio and video stream playing device |
CN115278249A (en) * | 2022-06-27 | 2022-11-01 | 北京大学 | Video block-level rate-distortion optimization method and system based on visual self-attention network |
WO2022257130A1 (en) * | 2021-06-11 | 2022-12-15 | Oppo广东移动通信有限公司 | Encoding method, decoding method, code stream, encoder, decoder, system and storage medium |
WO2022257049A1 (en) * | 2021-06-09 | 2022-12-15 | Oppo广东移动通信有限公司 | Encoding method, decoding method, code stream, encoder, decoder and storage medium |
CN116320410A (en) * | 2021-12-21 | 2023-06-23 | 腾讯科技(深圳)有限公司 | Data processing method, device, equipment and readable storage medium |
WO2023151365A1 (en) * | 2022-02-10 | 2023-08-17 | 腾讯科技(深圳)有限公司 | Image filtering method and apparatus, device, storage medium and program product |
WO2023221599A1 (en) * | 2022-05-18 | 2023-11-23 | 腾讯科技(深圳)有限公司 | Image filtering method and apparatus and device |
WO2023231775A1 (en) * | 2022-05-31 | 2023-12-07 | 华为技术有限公司 | Filtering method, filtering model training method and related device |
CN115278249B (en) * | 2022-06-27 | 2024-06-28 | 北京大学 | Video block-level rate distortion optimization method and system based on visual self-attention network |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107197260A (en) * | 2017-06-12 | 2017-09-22 | 清华大学深圳研究生院 | Video coding post-filter method based on convolutional neural networks |
WO2017222140A1 (en) * | 2016-06-24 | 2017-12-28 | 한국과학기술원 | Encoding and decoding methods and devices including cnn-based in-loop filter |
CN108134932A (en) * | 2018-01-11 | 2018-06-08 | 上海交通大学 | Filter achieving method and system in coding and decoding video loop based on convolutional neural networks |
CN108184129A (en) * | 2017-12-11 | 2018-06-19 | 北京大学 | A kind of video coding-decoding method, device and the neural network for image filtering |
CN108520505A (en) * | 2018-04-17 | 2018-09-11 | 上海交通大学 | Based on Multi net voting joint mapping and adaptively selected loop filtering implementation method |
-
2019
- 2019-06-13 CN CN201910511992.7A patent/CN110351568A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2017222140A1 (en) * | 2016-06-24 | 2017-12-28 | 한국과학기술원 | Encoding and decoding methods and devices including cnn-based in-loop filter |
CN107197260A (en) * | 2017-06-12 | 2017-09-22 | 清华大学深圳研究生院 | Video coding post-filter method based on convolutional neural networks |
CN108184129A (en) * | 2017-12-11 | 2018-06-19 | 北京大学 | A kind of video coding-decoding method, device and the neural network for image filtering |
CN108134932A (en) * | 2018-01-11 | 2018-06-08 | 上海交通大学 | Filter achieving method and system in coding and decoding video loop based on convolutional neural networks |
CN108520505A (en) * | 2018-04-17 | 2018-09-11 | 上海交通大学 | Based on Multi net voting joint mapping and adaptively selected loop filtering implementation method |
Non-Patent Citations (2)
Title |
---|
方涛: "《《中国优秀硕士学位论文全文数据库 信息科技辑》》", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
纪秋佳: "《基于半监督自步学习的跨任务深度网络应用于图像分类》", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110933422A (en) * | 2019-11-22 | 2020-03-27 | 南京信息工程大学 | HEVC loop filtering method based on EDCNN |
CN110933422B (en) * | 2019-11-22 | 2022-07-22 | 南京信息工程大学 | HEVC loop filtering method based on EDCNN |
CN113132729B (en) * | 2020-01-15 | 2023-01-13 | 北京大学 | Loop filtering method based on multiple reference frames and electronic device |
CN113132729A (en) * | 2020-01-15 | 2021-07-16 | 北京大学 | Loop filtering method based on multiple reference frames and electronic device |
CN113259671A (en) * | 2020-02-10 | 2021-08-13 | 腾讯科技(深圳)有限公司 | Loop filtering method, device and equipment in video coding and decoding and storage medium |
CN111405283A (en) * | 2020-02-20 | 2020-07-10 | 北京大学 | End-to-end video compression method, system and storage medium based on deep learning |
CN111405283B (en) * | 2020-02-20 | 2022-09-02 | 北京大学 | End-to-end video compression method, system and storage medium based on deep learning |
CN113452998A (en) * | 2020-03-25 | 2021-09-28 | 杭州海康威视数字技术股份有限公司 | Decoding, encoding and decoding method, device and equipment thereof |
CN114007082B (en) * | 2020-03-25 | 2022-12-23 | 杭州海康威视数字技术股份有限公司 | Decoding, encoding and decoding methods, devices and equipment |
CN114007082A (en) * | 2020-03-25 | 2022-02-01 | 杭州海康威视数字技术股份有限公司 | Decoding, encoding and decoding method, device and equipment thereof |
CN113452998B (en) * | 2020-03-25 | 2022-05-31 | 杭州海康威视数字技术股份有限公司 | Decoding, encoding and decoding method, device and equipment thereof |
CN111541894A (en) * | 2020-04-21 | 2020-08-14 | 电子科技大学 | Loop filtering method based on edge enhancement residual error network |
CN111866521B (en) * | 2020-07-09 | 2022-04-01 | 浙江工商大学 | Video image compression artifact removing method |
CN111866521A (en) * | 2020-07-09 | 2020-10-30 | 浙江工商大学 | Video image compression artifact removing method combining motion compensation and generation type countermeasure network |
CN113852850A (en) * | 2020-11-24 | 2021-12-28 | 广东朝歌智慧互联科技有限公司 | Audio and video stream playing device |
CN113852850B (en) * | 2020-11-24 | 2024-01-09 | 广东朝歌智慧互联科技有限公司 | Audio/video stream playing device |
CN113012073A (en) * | 2021-04-01 | 2021-06-22 | 清华大学 | Training method and device for video quality improvement model |
WO2022257049A1 (en) * | 2021-06-09 | 2022-12-15 | Oppo广东移动通信有限公司 | Encoding method, decoding method, code stream, encoder, decoder and storage medium |
WO2022257130A1 (en) * | 2021-06-11 | 2022-12-15 | Oppo广东移动通信有限公司 | Encoding method, decoding method, code stream, encoder, decoder, system and storage medium |
CN113613000A (en) * | 2021-08-20 | 2021-11-05 | 天津大学 | Intelligent multi-resolution depth video intra-frame prediction method |
CN113613000B (en) * | 2021-08-20 | 2024-04-26 | 天津大学 | Intelligent multi-resolution depth video intra-frame prediction method |
CN116320410A (en) * | 2021-12-21 | 2023-06-23 | 腾讯科技(深圳)有限公司 | Data processing method, device, equipment and readable storage medium |
WO2023116173A1 (en) * | 2021-12-21 | 2023-06-29 | 腾讯科技(深圳)有限公司 | Data processing method and apparatus, and computer device and storage medium |
WO2023151365A1 (en) * | 2022-02-10 | 2023-08-17 | 腾讯科技(深圳)有限公司 | Image filtering method and apparatus, device, storage medium and program product |
WO2023221599A1 (en) * | 2022-05-18 | 2023-11-23 | 腾讯科技(深圳)有限公司 | Image filtering method and apparatus and device |
WO2023231775A1 (en) * | 2022-05-31 | 2023-12-07 | 华为技术有限公司 | Filtering method, filtering model training method and related device |
CN115278249A (en) * | 2022-06-27 | 2022-11-01 | 北京大学 | Video block-level rate-distortion optimization method and system based on visual self-attention network |
CN115278249B (en) * | 2022-06-27 | 2024-06-28 | 北京大学 | Video block-level rate distortion optimization method and system based on visual self-attention network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110351568A (en) | A kind of filtering video loop device based on depth convolutional network | |
CN107018422B (en) | Still image compression method based on depth convolutional neural networks | |
CN107197260B (en) | Video coding post-filter method based on convolutional neural networks | |
CN112203093B (en) | Signal processing method based on deep neural network | |
CN108921910B (en) | JPEG coding compressed image restoration method based on scalable convolutional neural network | |
CN108184129A (en) | A kind of video coding-decoding method, device and the neural network for image filtering | |
CN105430416B (en) | A kind of Method of Fingerprint Image Compression based on adaptive sparse domain coding | |
CN107463989A (en) | A kind of image based on deep learning goes compression artefacts method | |
CN108900848A (en) | A kind of video quality Enhancement Method based on adaptive separable convolution | |
CN110136057B (en) | Image super-resolution reconstruction method and device and electronic equipment | |
CN111885280B (en) | Hybrid convolutional neural network video coding loop filtering method | |
CN107181949A (en) | A kind of compression of images framework of combination super-resolution and residual coding technology | |
CN107645662A (en) | A kind of colour image compression method | |
CN107734333A (en) | A kind of method for improving video error concealing effect using network is generated | |
CN103546759A (en) | Image compression coding method based on combination of wavelet packets and vector quantization | |
CN109978772A (en) | Based on the deep learning compression image recovery method complementary with dual domain | |
CN113132729B (en) | Loop filtering method based on multiple reference frames and electronic device | |
CN112019854B (en) | Loop filtering method based on deep learning neural network | |
CN112218094A (en) | JPEG image decompression effect removing method based on DCT coefficient prediction | |
CN112188217B (en) | JPEG compressed image decompression effect removing method combining DCT domain and pixel domain learning | |
CN110223224A (en) | A kind of Image Super-resolution realization algorithm based on information filtering network | |
CN116347107A (en) | QP self-adaptive loop filtering method based on variable CNN for VVC video coding standard | |
CN1825894A (en) | All phase cosine double orthogonal transformation and JPEG improving method | |
CN114596223A (en) | Quantization parameter self-adaptive convolution neural network loop filter and construction method thereof | |
CN110148087B (en) | Image compression and reconstruction method based on sparse representation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191018 |