CN108921910B - JPEG coding compressed image restoration method based on scalable convolutional neural network - Google Patents
JPEG coding compressed image restoration method based on scalable convolutional neural network Download PDFInfo
- Publication number
- CN108921910B CN108921910B CN201810853338.XA CN201810853338A CN108921910B CN 108921910 B CN108921910 B CN 108921910B CN 201810853338 A CN201810853338 A CN 201810853338A CN 108921910 B CN108921910 B CN 108921910B
- Authority
- CN
- China
- Prior art keywords
- image
- neural network
- convolutional neural
- scalable
- image feature
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
- Compression Of Band Width Or Redundancy In Fax (AREA)
Abstract
The invention discloses a JPEG coding compressed image restoration method based on a scalable convolutional neural network, and belongs to the field of image signal processing and artificial intelligence. The invention firstly provides a processing frame of a JPEG coding compressed image restoration method based on a scalable convolutional neural network, which consists of an image feature coding unit, an image feature decoding unit, a nonlinear feature mapping unit and a greedy loss frame for training the scalable convolutional neural network. The method utilizes the local optimal characteristic of the greedy loss frame and combines the strong generalization capability of the deep convolutional neural network, thereby realizing the dynamic adjustment of the depth of the network on the premise of not changing the weight of the network model. Meanwhile, compared with the traditional image deblocking filtering method in the field of JPEG coding compressed image restoration, the method has obvious improvement on subjective image quality and key technical indexes such as PSNR (picture support ratio), SSIM (structural similarity) and the like.
Description
Technical Field
The invention belongs to the field of image signal processing and artificial intelligence, and particularly relates to a JPEG (joint photographic experts group) coding compressed image restoration method.
Background
With the rapid development and wide application of multimedia technology, high-quality images and videos have become a mainstream demand. The higher the quality of the video data, the greater its analytical value. However, both images and video have been transmitted in compressed form and stored in a normal state, subject to limited channel bandwidth and storage capacity. The most common image degradation factors are down-sampling and compression distortion. The down-sampling reduces the spatial resolution of the image, and compression distortion causes the image to have the problems of blocking effect, ringing, blurring and the like. Therefore, the method has important theoretical significance and practical application value in researching the multi-degradation factor image restoration technology aiming at the low-resolution image with compression distortion.
The image super-resolution restoration method can restore a high-resolution image by using a single-frame or multi-frame low-resolution image. When the super-resolution method aiming at the uncompressed image is adopted to directly carry out super-resolution restoration on the compressed and distorted low-resolution image, the image resolution can be improved, but the serious block effect distortion phenomenon can be amplified.
The JPEG image compression coding technology adopts DCT transformation based on 8 x 8 pixel blocks, and the transformed DCT coefficients are quantized to eliminate redundant information of an image space, so that the original image is compressed and stored. Due to the excellent compression performance and low computational complexity of JPEG, this method is widely applied to various fields related to image technology, and is one of the most widely applied image compression and encoding technologies in the world.
But since JPEG is a lossy compression method, a higher compression ratio is achieved by using a larger quantization step, which causes a more serious degradation of image quality while obtaining a higher compression ratio. Since the quantization operation is a non-linear operation and all pixel blocks in the image are quantized independently, the boundary of adjacent pixel blocks generates obvious gray jump and block effect, and a large amount of disordered fuzzy and false edges (ringing effect) are formed in the pixel blocks. This can have a significant negative impact on either the viewing experience of the image or the application of image-based computer vision techniques.
On the other hand, with the rapid development of the computing power of modern processors and the rapid maturation of deep learning theory research, a method based in part on a deep convolutional neural network has been applied to the restoration of JPEG compression-encoded images. Although the method can achieve a good recovery effect, the method is often large in calculation amount, needs to occupy fixed and large memory and calculation resources, and is difficult to perform real-time task scheduling in a multitask system, so that the application of the method in an actual engineering project is greatly limited.
Disclosure of Invention
In order to solve the problems, the invention provides a method for restoring a JPEG (joint photographic experts group) compression coded image based on a scalable convolutional neural network, so as to improve the visual effect of the image, and meanwhile, the constructed scalable convolutional neural network framework allows the convolutional neural network to dynamically adjust the depth of the network in the working process so as to adapt to the real-time scheduling of a multitask system.
The technical scheme provided by the invention is as follows:
a method for restoring JPEG compressed and coded images based on a scalable convolutional neural network comprises the following steps:
step 2, constructing a scalable convolutional neural network model, where the scalable convolutional neural network model includes an image feature coding unit, a mapping group including N nonlinear feature mapping units connected in sequence, and an image feature decoding unit, where the image feature coding unit codes features of a received color image, outputs an image feature code to the mapping group, the mapping group selects m times of mapping processing on the received image feature code by using m adjacent nonlinear feature mapping units, outputs the mapped image feature code to an image feature decoding unit, and the image feature decoding unit decodes the received image feature code and outputs the decoded color image, where m is equal to or less than N, N is 1,2,3, …, and N is a natural number;
step 3, constructing a greedy loss frame based on the scalable convolutional neural network model, wherein the greedy loss frame is as follows: on the basis of a scalable convolutional neural network model, an image feature decoding unit is connected behind each mapping nonlinear feature mapping unit, and a loss calculation unit is connected behind each image feature decoding unit;
step 4, training the constructed greedy loss frame by using a training set, and determining a network weight parameter after the training is finished;
and 5, during application, inputting the compressed image to be restored into the trained scalable convolutional neural network model, sequentially processing the compressed image to be restored by the image feature coding unit, the m nonlinear feature mapping units selected according to the system operation capacity and the image feature decoding unit, and outputting a restored image.
According to the restoration method provided by the invention, the local optimal characteristic of a greedy loss frame is utilized, and the strong generalization capability of the deep convolutional neural network is combined, so that the depth of the network is dynamically adjusted on the premise of not changing the weight of a network model, and the requirement of not passing through the deep convolutional neural network can be met.
Preferably, when dividing, the overlapping rate between the adjacent image blocks is allowed to be not more than 50%.
Preferably, the size of the divided image block is 48 × 48 pixels, and the division should ensure that the horizontal and vertical coordinates of the upper left corner point of the image block in the original high-definition image and the distorted image are not integer multiples of 8.
The image feature coding unit comprises a convolutional layer CONV _ E1, an active layer RELU _ E1, a convolutional layer CONV _ E2 and an active layer RELU _ E2 which are connected in sequence.
The nonlinear feature mapping unit comprises a convolution layer CONV _ M1, an activation layer RELU _ M1, a convolution layer CONV _ M2 and an amplification layer AMP _ M1 which are connected in sequence, and further comprises a fusion layer CONCAT _ M1 which fuses the output of the square layer AMP _ M1 and the input of the convolution layer CONV _ M1.
The image feature decoding unit comprises a convolutional layer CONV _ D1, an active layer RELU _ D1, a convolutional layer CONV _ D2 and an active layer RELU _ D2 which are connected in sequence.
The convolution and size of all convolutional layers described above are the same as the convolution mode: the size of the convolution kernel is 3 multiplied by 3, the number of the convolution kernels is 256, the sliding step length is 1, and the edge filling is 1;
the activation functions of all activation layers are ReLU functions:
the magnification factor λ is 0.1 for all the magnification layers.
Preferably, the mapping set includes 8 nonlinear feature mapping units connected in sequence. The constructed scalable convolutional neural network model comprises an image feature coding layer ENCODER, nonlinear feature mapping units MAPPER _1 and MAPPER _2 … MAPPER _8 and an image feature decoding layer DECODER which are connected in sequence.
Before a greedy loss frame is trained, an Adam optimizer is adopted in the learning optimization method for setting the scalable convolutional neural network model, the initial learning rate is set to be 0.0001, and the attenuation of the learning rate is realized every 20000 times
Until the learning rate is less than 2 x 10-6Then no attenuation is occurring and the maximum number of iterations is set to 2 x 105。
In the training process, after the distorted image is processed by the image feature coding unit, the mapping group and the image feature decoding unit in sequence, the loss of each loss calculation unit is calculated, each loss is weighted and superposed to obtain the final loss, and then the final loss is used for carrying out anti-echo propagation to update the network weight parameters.
The method for restoring the JPEG compressed and coded image based on the scalable convolutional neural network has the advantages that:
the networks with different depths do not need to be trained for the image with the same JPEG encoding quality, and all the networks with the depths not exceeding the depth of the current network can be obtained only by training the network with the maximum depth. Compared with a common method for training a plurality of networks with different depths, the scalable convolutional network constructed based on the greedy loss framework can greatly reduce the training time of a network model and can quickly evaluate the influence of the network depth on the network performance. Meanwhile, compared with the situation that a large amount of memory space and disk I/O (input/output) are occupied when a plurality of network models are loaded in a real-time multitask system, the scalable convolutional neural network only needs to be loaded once, and the scheduling efficiency of the system can be effectively improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
Fig. 1 is a block flow diagram of a method for recovering a JPEG-encoded compressed image based on a scalable convolutional neural network according to an embodiment;
FIG. 2 is a schematic structural diagram of a baseline model of a scalable convolutional neural network provided by an embodiment;
FIG. 3 is a schematic diagram of the structure of the image feature encoding unit in FIG. 2;
FIG. 4 is a schematic diagram of the structure of the image feature decoding unit in FIG. 2;
FIG. 5 is a schematic diagram of the structure of the nonlinear feature mapping unit in FIG. 2;
fig. 6 is an evaluation of the computational performance of the scalable convolutional neural network of the present invention when using different depths.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
As shown in fig. 1, the method for restoring a JPEG compressed and encoded image based on a scalable convolutional neural network according to this embodiment includes the following steps:
step 1: constructing training data
The dataset used to train the scalable convolutional neural network model proposed by the present invention is the DIV2K dataset. The DIV2K data set contains 900 high-definition color images at 2K resolution. Firstly, 900 images are coded according to the JPEG compression coding quality of a restoration target, and 900 images which are subjected to JPEG compression coding are obtained. The images were divided into two groups, one of which was a training data set containing 800 images and the other was a validation data set containing 100 images. For the training data set, the images in the training data set are cut into a plurality of 48 × 48 image blocks according to the random segmentation step of [37,61 ]. For the verification dataset, the original image is divided into 16 sub-images of 4 × 4 on average, and one image block of 48 × 48 is randomly scratched in each sub-image, resulting in a total of 16 × 100 to 1600 image blocks.
Step 2: building scalable convolutional neural network model
By using a deep learning framework Keras and a rear-end framework Tensorflow, a scalable convolutional neural network model provided by the invention is built, as shown in FIG. 2. The base line of the model consists of an image characteristic coding unit, an image characteristic decoding unit and 8 nonlinear characteristic mapping units.
As shown in fig. 3, the image feature encoding unit includes convolutional layers CONV _ E1 and CONV _ E2, and active layers RELU _ E1 and RELU _ E2. As shown in fig. 4, the image feature decoding unit is composed of convolutional layers CONV _ D1, CONV _ D2, active layers RELU _ D1, and RELU _ D2. As shown in fig. 5, the nonlinear feature mapping unit is composed of volume base layers CONV _ M1, CONV _ M2, an activation layer RELU _ M1, an amplification layer AMP _ M1, and a fusion layer CONCAT _ M1.
All convolutional layers except convolutional layer CONV _ D2 contain 256 convolutional filters, while convolutional layer CONV _ D2 contains 3 convolutional filters. The convolution kernel size of all convolutional layers except CONV _ D2 and CONV _ E1 was 3 × 3, and the convolution kernel size of CONV _ D2 and CONV _ E1 was 5 × 5.
The magnification of the magnifying layer in all non-linear feature cells is 0.1.
The activation functions of all the activation layers are ReLU functions.
And step 3: training scalable convolutional neural networks
When training the network model, the size of the input layer is 48 × 48 × 3; when the image is restored, the size of the input layer is the actual size of the image to be restored. If the image to be restored is a gray-scale image, copying 2 parts of the gray-scale image, combining the gray-scale image and the original image into a 3-channel image, and inputting the 3-channel image into the network model.
When a network model is trained, a greedy Loss frame is introduced, and the same image feature decoding unit and an L2Loss layer are connected after the output of each nonlinear decoding unit, so that 8L 2Loss layers are calculated. Each lossy layer has the same loss weight and the sum of these loss weights is 1, i.e. the weight of each lossy layer is 0.125.
An Adam optimizer is adopted during model training, the initial learning rate is set to be 0.0001, and the attenuation of the learning rate is realized every 20000 times of iteration
Until the learning rate is less than 2 x 10-6Then there is no attenuation of the light beam,the maximum number of iterations is set to 2 × 105。
Each batch of training data comprises 32 48 × 48 × 3 image blocks. The training data is propagated forward and the weighted Loss of each L2Loss is calculated, the Loss is propagated backward and the model parameters are updated. And repeating the forward propagation and the backward propagation until the model converges. During model training, validation losses were calculated on the validation set after each 5000 batches of data were trained. The condition for considering the convergence of the model is that after the learning rate is not attenuated any more, the loss value of the verification set does not decrease for 3 consecutive times or the training batch reaches the maximum iteration number.
And 4, step 4: restoration of JPEG compression coded images using trained models
A baseline model of the scalable neural network is first built and the trained model is loaded. The image feature decoding unit is connected to the nonlinear feature mapping unit of the target depth. And inputting the JPEG compressed and coded image into a network to obtain a restored image.
The model provided by the invention is evaluated by using a BSDS500 data set:
the indicators used for the evaluation are peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM). The larger the peak signal-to-noise ratio and the structural similarity, the better. The evaluation results are shown in table 1 below, and it can be seen that the convolutional neural network model provided by the present invention has superior performance.
Further, the computational performance of the scalable convolutional neural network of the present invention when using different depths was evaluated on NVIDIA K80 GPU server, using an index of the number of pixels processed per second (MCP/S). The larger the index is, the faster the calculation speed is. The evaluation results are shown in fig. 6.
TABLE 1
Wherein ARCNN is a restoration method in the document "C.Dong, Y.Deng, C.L.Chen et al: Compression architecture by a Deep conditional network. IEEE Conference on computer Vision. IEEE,576-584 (2016)"; TNRD is a restoration method in the literature "Chen, Y., Yu, W., pack, T." Online optimized differentiation processes for effective imaging. in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.5261-5269 (2015) "; DDCN is a restoration method in the document "J.Guo, and H.Chao: Building Dual-Domain retrieval for Compression ArtifactsRegulation. European Conference on Computer Vision. Springer, Cham,628-644 (2016)".
In summary, the scalable convolutional neural network provided by the present invention not only greatly improves the computation speed and the image restoration result compared with the previous method, but also dynamically adjusts the depth of the network during operation, and can well achieve the dynamic balance between the image restoration quality and the computation efficiency.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.
Claims (9)
1. A method for restoring JPEG compressed and coded images based on a scalable convolutional neural network comprises the following steps:
step 1, JPEG compression coding is carried out on the high-definition image by adopting a fixed image quality factor to obtain a distorted image, and the high-definition image and the distorted image are divided into a plurality of groups of image blocks by adopting the same random step length to form a training set;
step 2, constructing a scalable convolutional neural network model, where the scalable convolutional neural network model includes an image feature coding unit, a mapping group including N nonlinear feature mapping units connected in sequence, and an image feature decoding unit, where the image feature coding unit codes features of a received color image, outputs an image feature code to the mapping group, the mapping group selects m times of mapping processing on the received image feature code by using m adjacent nonlinear feature mapping units, outputs the mapped image feature code to an image feature decoding unit, and the image feature decoding unit decodes the received image feature code and outputs the decoded color image, where m is equal to or less than N, N is 1,2,3, …, and N is a natural number;
the nonlinear feature mapping unit comprises a convolution layer CONV _ M1, an activation layer RELU _ M1, a convolution layer CONV _ M2 and an amplification layer AMP _ M1 which are connected in sequence, and further comprises a fusion layer CONCAT _ M1 for fusing the output of the amplification layer AMP _ M1 and the input of the convolution layer CONV _ M1;
step 3, constructing a greedy loss frame based on the scalable convolutional neural network model, wherein the greedy loss frame is as follows: on the basis of a scalable convolutional neural network model, an image feature decoding unit is connected behind each nonlinear feature mapping unit, and a loss calculation unit is connected behind each image feature decoding unit;
step 4, training the constructed greedy loss frame by using a training set, and determining a network weight parameter after the training is finished;
and 5, during application, inputting the compressed image to be restored into the trained scalable convolutional neural network model, sequentially processing the compressed image to be restored by the image feature coding unit, the m nonlinear feature mapping units selected according to the system operation capacity and the image feature decoding unit, and outputting a restored image.
2. The method of claim 1, wherein the segmentation allows no more than 50% overlap between adjacent image blocks.
3. The method for restoring the JPEG-compressed coded image based on scalable convolutional neural network as claimed in claim 1 or 2, wherein the size of the segmented image block is 48 × 48 pixels, and the segmentation should ensure that the horizontal and vertical coordinates of the upper left corner point of the image block in the original high-definition image and the distorted image are not integer multiples of 8.
4. The scalable convolutional neural network-based JPEG compression encoded image restoration method as claimed in claim 1, wherein the image feature encoding unit includes a convolutional layer CONV _ E1, an active layer RELU _ E1, a convolutional layer CONV _ E2, and an active layer RELU _ E2, which are connected in this order.
5. The scalable convolutional neural network-based JPEG compression-encoded image restoration method as claimed in claim 4, wherein said image feature decoding unit comprises a convolutional layer CONV _ D1, an activation layer RELU _ D1, a convolutional layer CONV _ D2, and an activation layer RELU _ D2, which are connected in sequence.
6. The scalable convolutional neural network-based JPEG compression encoded image restoration method as claimed in claim 1, wherein the mapping group includes 8 nonlinear feature mapping units connected in sequence.
7. The scalable convolutional neural network-based JPEG compression-encoded image restoration method according to claim 5, wherein convolution and size of all convolutional layers are identical in convolution mode: the size of the convolution kernel is 3 multiplied by 3, the number of the convolution kernels is 256, the sliding step length is 1, and the edge filling is 1;
the activation functions of all activation layers are ReLU functions:
the magnification factor λ is 0.1 for all the magnification layers.
8. The method for restoring a JPEG compression-coded image based on a scalable convolutional neural network as claimed in claim 5, wherein before training the greedy loss framework, the learning optimization method for setting the scalable convolutional neural network model adopts an Adam optimizer, the initial learning rate is set to 0.0001, and the attenuation of the learning rate is performed every 20000 times
Until the learning rate is less than 2 x 10-6Then no further attenuation is performed and the maximum number of iterations is set to 2 x 105。
9. The method for restoring the JPEG compressed and encoded image based on the scalable convolutional neural network as claimed in claim 7, wherein in the training process, the distorted image is processed by the image feature encoding unit, the mapping group and the image feature decoding unit in sequence, the loss of each loss calculating unit is calculated, each loss is weighted and overlapped to obtain the final loss, and then the final loss is used for updating the network weight parameters in an echo propagation mode.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810853338.XA CN108921910B (en) | 2018-07-30 | 2018-07-30 | JPEG coding compressed image restoration method based on scalable convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810853338.XA CN108921910B (en) | 2018-07-30 | 2018-07-30 | JPEG coding compressed image restoration method based on scalable convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108921910A CN108921910A (en) | 2018-11-30 |
| CN108921910B true CN108921910B (en) | 2020-05-29 |
Family
ID=64417675
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810853338.XA Active CN108921910B (en) | 2018-07-30 | 2018-07-30 | JPEG coding compressed image restoration method based on scalable convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921910B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109903351B (en) * | 2019-03-02 | 2023-05-30 | 复旦大学 | Image compression method based on combination of convolutional neural network and traditional coding |
| CN110062125A (en) * | 2019-04-18 | 2019-07-26 | 西安万像电子科技有限公司 | Image transfer method and device |
| CN110223273B (en) * | 2019-05-16 | 2023-04-07 | 天津大学 | Image restoration evidence obtaining method combining discrete cosine transform and neural network |
| CN110519595A (en) * | 2019-08-08 | 2019-11-29 | 浙江大学 | A kind of jpeg compressed image restored method based on frequency domain quantization estimated amount of damage |
| CN112784952B (en) * | 2019-11-04 | 2024-03-19 | 珠海格力电器股份有限公司 | Convolutional neural network operation system, method and equipment |
| CN111010566A (en) * | 2019-12-04 | 2020-04-14 | 杭州皮克皮克科技有限公司 | Non-local network-based video compression distortion restoration method and system |
| CN113284202B (en) * | 2021-06-11 | 2022-10-25 | 北京大学深圳研究生院 | Image compression sensing method of scalable network based on content self-adaption |
| CN113674155A (en) * | 2021-08-25 | 2021-11-19 | 中国铁塔股份有限公司湖北省分公司 | Image super-resolution method, device and storage medium based on information aggregation network |
| CN115052148B (en) * | 2022-07-21 | 2023-04-25 | 南昌工程学院 | Image compression algorithm based on model segmentation compression self-encoder |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709875B (en) * | 2016-12-30 | 2020-02-18 | 北京工业大学 | Compressed low-resolution image restoration method based on joint depth network |
| CN108109121A (en) * | 2017-12-18 | 2018-06-01 | 深圳市唯特视科技有限公司 | A kind of face based on convolutional neural networks obscures quick removing method |
-
2018
- 2018-07-30 CN CN201810853338.XA patent/CN108921910B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN108921910A (en) | 2018-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921910B (en) | JPEG coding compressed image restoration method based on scalable convolutional neural network | |
| CN110363716B (en) | High-quality reconstruction method for generating confrontation network composite degraded image based on conditions | |
| CN111028150B (en) | Rapid space-time residual attention video super-resolution reconstruction method | |
| US11922599B2 (en) | Video super-resolution processing method and apparatus | |
| CN112801901A (en) | Image deblurring algorithm based on block multi-scale convolution neural network | |
| CN110351568A (en) | A kind of filtering video loop device based on depth convolutional network | |
| CN110349087B (en) | RGB-D image high-quality grid generation method based on adaptive convolution | |
| Hu et al. | An adaptive two-layer light field compression scheme using GNN-based reconstruction | |
| CN115661403A (en) | Explicit radiation field processing method, device and storage medium | |
| CN116542889A (en) | Panoramic video enhancement method with stable view point | |
| CN116630131A (en) | Coding and decoding system and method for invisible screen watermark | |
| CN114022356A (en) | River course flow water level remote sensing image super-resolution method and system based on wavelet domain | |
| Jin et al. | Post-processing for intra coding through perceptual adversarial learning and progressive refinement | |
| CN106961605B (en) | One kind being based on the matched light field image compression method of macro pixel boundary | |
| CN110677644A (en) | Video coding and decoding method and video coding intra-frame predictor | |
| CN117560511A (en) | Spacer image compression method and system based on graph segmentation technology and electric power inspection | |
| CN113362239A (en) | Deep learning image restoration method based on feature interaction | |
| CN115499666B (en) | Video compression method, video decompression method, video compression device, video decompression device, and storage medium | |
| CN113949883A (en) | Coding method, decoding method and coding and decoding system of bidirectional prediction frame | |
| Yang et al. | Graph-convolution network for image compression | |
| CN112468826B (en) | VVC loop filtering method and system based on multilayer GAN | |
| CN112261415B (en) | Image compression coding method based on overfitting convolution self-coding network | |
| CN114494472A (en) | Image compression method based on depth self-attention transformation network | |
| WO2022257674A1 (en) | Encoding method and apparatus using inter-frame prediction, device, and readable storage medium | |
| CN116760983B (en) | Loop filtering method and device for video coding |
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 |