CN109377532A - Image processing method and device neural network based - Google Patents
Image processing method and device neural network based Download PDFInfo
- Publication number
- CN109377532A CN109377532A CN201811212227.7A CN201811212227A CN109377532A CN 109377532 A CN109377532 A CN 109377532A CN 201811212227 A CN201811212227 A CN 201811212227A CN 109377532 A CN109377532 A CN 109377532A
- Authority
- CN
- China
- Prior art keywords
- image
- model
- comparing result
- loss
- characteristic pattern
- 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
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
-
- 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
- H04N19/85—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using pre-processing or post-processing specially adapted for video compression
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of image processing methods neural network based, it include: that operation is filled to the first image using generation model to obtain the second image, wherein, first image is the image for generate after size compression to original image, and the size of second image and the size of the original image are at designated ratio；Supervised learning is carried out to original image and the second image using discrimination model to obtain the first comparing result, and details of use contrast model calculates the characteristic pattern of original image and the characteristic pattern of the second image, and then determine the second comparing result, wherein, the second comparing result indicates the difference of the characteristic pattern of original image and the characteristic pattern of the second image；It is trained based on the first comparing result and the second comparing result and generates model and discrimination model.Image processing method of the invention can restore the image of higher compression ratios, to increase substantially the compression ratio of image, save the bandwidth demand and memory space requirements of transmission images serve.
Description
Technical field
The present invention relates to image procossing more particularly to image processing methods neural network based and device.
Background technique
Compression of images refers to the technology for damaging with less data amount or nondestructively indicating original picture element matrix, thus with more
Efficient format is added to store and transmit data.Why image data can be compressed, be exactly because there is redundancy in data, than
As: caused by the correlation of spatial redundancy caused by the correlation in image between adjacent pixel, different color planes or spectral band
Spectral redundancy etc..The purpose of data compression is exactly bit number needed for reducing expression data by removing these data redundancies.
The compression of image is divided into lossy compression and lossless compression, and lossy compression refers to neglects some insensitive letters in compression process
Breath, although former data cannot be restored completely, can obtain higher compression ratio.Lossless compression then refers to the redundancy for only removing data
Information is recompiled by way of comentropy, to obtain compression ratio as high as possible, the data energy after lossless compression
It is enough completely to restore former data.
Image amplification refers to the amplification carried out in size to image.Obviously the information in original image is not enough to comprising putting
All image informations after big, so this amplification must be distortion.Image multiplication method is studied and how to be reduced to the greatest extent
In the case where distortion, by different interpolation methods, original image is amplified in specified size.Common image, which amplifies, to be calculated
Method has the methods of bilinear interpolation, bi-cubic interpolation.
Neural network (Neural Network) is a kind of network structure artificially designed, and essence is multi-layer perception (MLP)
(Multi-layer Perceptron).Perceptron is made of several neurons (Neuron), each neuron from it is external or its
Its node receives input signal, and obtains output signal by activation primitive, just as the signal transmitting of neuron in brain.Nerve
Member is connected by layer, forms network structure.It is different from nerve cell, the signal of artificial neuron can backpropagation, this feedback
Mechanism allows perceptron to have learning functionality.In addition to learning functionality, multi-layer perception (MLP) can indicate Nonlinear Mapping, therefore nerve net
It is some relative complex that network can help people to solve the problems, such as, such as pattern-recognition, automatic control, Decision Evaluation, prediction.Using
Scene is divided into supervised learning, unsupervised learning and semi-supervised learning etc., described by defining different loss functions target,
Fitting parameter achievees the purpose that network.
Summary of the invention
Aiming at the problem that can only restore the image compared with low compression ratio in image procossing at present, one aspect of the present invention is provided
A kind of image processing method neural network based, comprising: operation is filled to the first image using model is generated
Obtain the second image, wherein the first image is the image for generate after size compression to original image, second figure
The size of picture and the size of the original image are at designated ratio；Using discrimination model to the original image and second figure
The first comparing result is obtained as carrying out supervised learning, and details of use contrast model calculates the spy of the original image
The characteristic pattern of sign figure and second image, and then determine the second comparing result, wherein described in the second comparing result expression
The difference of the characteristic pattern of the characteristic pattern of original image and second image；Based on first comparing result and second pair described
The generation model and the discrimination model are trained than result.
In one embodiment, the method also includes: third image is restored using the generation model after training, wherein
The third image is compressed image.
In one embodiment, the size of second image is identical as the size of the original image.
In one embodiment, operation is filled to the first image come to obtain the second image include: to make using model is generated
Multilayer convolution algorithm is carried out to the first image with generation model to obtain the second image.
In one embodiment, the generation model is trained based on first comparing result and second comparing result
It include: to be lost using first comparing result and second comparing result as model with the discrimination model, using reversed
Transmission method trains the generation model and the discrimination model.
In one embodiment, first comparing result is loss of correcting errors, and second comparing result is details damage
It loses.
In one embodiment, the loss of detail is the loss of detail for the key area being labeled.
In one embodiment, the Detail contrast model is constructed based on the characteristic pattern hidden layer of Detail contrast network.
In one embodiment, the generation model and/or the judgment models are constructed based on convolutional neural networks.
Another aspect of the present invention provides a kind of image processing apparatus neural network based, comprising: memory is used for
Store instruction；And processor, it is coupled to the memory, described instruction makes the dress when being executed by the processor
It sets and executes method described in any of the above embodiments.
Another aspect of the present invention provides a kind of computer readable storage medium, and the storage medium includes instruction, described
Instruction is performed, so that the processor of the computer executes method described in any of the above embodiments.
Image processing method of the invention can restore the image of higher compression ratios, to increase substantially the compression of image
Than saving the bandwidth demand and memory space requirements of transmission images serve；Compared with traditional image interpolation interpolator arithmetic, pass through
Convolutional neural networks carry out image generation, and obtained image is more clear, precisely, and details restores more preferably, can carry higher
Magnification ratio；Reinforce generating the image detail reduction degree of model as additional loss by introduced feature picture scroll lamination；It can
Enhance model performance to define by the loss function of more customizedization, such as: to the image data base of file photocopy
The character area of sample is labeled, and is increased the loss of detail punishment of character area, is kept generation model quicker to character area
Sense, protrudes prior information according to demand.
Detailed description of the invention
The flow chart 100 of the image processing method neural network based of Fig. 1 embodiment according to the present invention；
The flow chart 200 of one embodiment of Fig. 2 image processing method neural network based according to the present invention；
The schematic diagram of the image processing apparatus neural network based 300 of Fig. 3 the embodiment of the present invention.
Specific embodiment
Below with reference to each exemplary embodiment of the attached drawing detailed description disclosure.Flow chart and block diagram in attached drawing are shown
The architecture, function and operation in the cards of method and system according to various embodiments of the present disclosure.It should be noted that
Each box in flowchart or block diagram can represent a part of a module, program segment or code, the module, program
Section or a part of code may include one or more holding for realizing the logic function of defined in each embodiment
Row instruction.It should also be noted that in some alternative implementations, function marked in the box can also be attached according to being different from
The sequence marked in figure occurs.For example, two boxes succeedingly indicated can actually be basically executed in parallel or it
Can also execute in a reverse order sometimes, this depend on related function.It should also be noted that flow chart
And/or the combination of each box in block diagram and the box in flowchart and or block diagram, function as defined in execution can be used
Can or the dedicated hardware based system of operation realize, or specialized hardware can be used and the combination of computer instruction comes
It realizes.
Term as used herein "include", "comprise" and similar terms are understood to open term, i.e.,
" including/including but not limited to ", expression can also include other content.Term "based" is " being based at least partially on ".Term
" one embodiment " expression " at least one embodiment "；Term " another embodiment " expression " at least one other embodiment ",
Etc..
Technology, method and apparatus known to person of ordinary skill in the relevant may be not discussed in detail, but suitable
In the case of, the technology, method and apparatus should be considered as part of specification.For between each unit in attached drawing
Line, it is only for convenient for explanation, indicate that the unit at least line both ends is in communication with each other, it is not intended that limitation does not connect
It can not be communicated between the unit of line.
Term is explained
Correct errors loss: describe the image generated with original image compared to whether really quantization measurement, for example, using cross entropy
To measure.
Loss of detail: the quantization of the difference between two image details of description is measured.
Detail contrast network: contrast images details obtains the neural network of loss of detail, i.e., classifies by large-scale image
The network that several convolutional layers for the deep neural network that data set (for example, ImageNet data set) training is completed are formed, for example,
Residual error network (ResNet).
Fig. 1 shows the flow chart 100 of the image processing method neural network based of embodiment according to the present invention.
Step S101: operation is filled to the first image using generation model to obtain the second image, wherein the first figure
As the image generate after size compression to original image, the size of second image and the size of the original image
At designated ratio.Original image referred to herein includes uncompressed scene image, such as user identity card, passport, text
Images of scenes such as part photocopy etc..
Step S102: supervised learning is carried out to original image and the second image using discrimination model to obtain the first comparison
As a result, and details of use contrast model calculate the characteristic pattern of original image and the characteristic pattern of the second image, and then determine the
Two comparing results, wherein the second comparing result indicates the difference of the characteristic pattern of original image and the characteristic pattern of the second image.One
In kind embodiment, Detail contrast model can be constructed based on the characteristic pattern hidden layer of Detail contrast network.Another real
It applies in mode, Detail contrast model can be constructed based on fully-connected network, and be instructed using large-scale image categorized data set
Practice.
Step S103: it is trained based on the first comparing result and the second comparing result and generates model and discrimination model.
It should be understood that the generation model and/or discrimination model in the present invention are constructed based on deep neural network.One
In a embodiment, generates model and/or discrimination model is constructed based on convolutional neural networks.In another embodiment, it generates
Model and/or discrimination model are constructed based on fully-connected network.
The flow chart 200 of one embodiment of Fig. 2 image processing method neural network based according to the present invention.
Step S201: operation is filled to the first image by generating model and obtains the original of size and the first image
Identical second image of picture size, wherein the first image is that size compression is carried out to original image (for example, length and width respectively compress 2
The image generated afterwards again, it is preferable that length and width respectively compress 4 times, and the multiple compressed certainly can also be higher).Here the first image
It can be obtained from corresponding image data base with its original image.In one embodiment, model is generated by multilayer deconvolution
The input of first image is generated in model, generates model Output Size identical with original image size the by neural network composition
Two images.It should be understood that residual block (residual block), crowd regularization (batch can be used in generating model
) etc. normalization modules improve the ability that model generates image.It should be understood that the first image can for by pair
Original image carries out the image that other suitable compress modes (for example, various damage or lossless compress mode) generate, and
The size of second image can be with the size of original image at any suitable ratio.
Step S202: discrimination model carries out supervised learning to the original image of the first image and the second image to obtain just
It accidentally loses, and the characteristic pattern hidden layer of details of use comparison network calculates the characteristic pattern and second image of the original image
Characteristic pattern, and then determine loss of detail, wherein loss of detail indicates the characteristic pattern of the original image and the characteristic pattern of the second image
Difference.It should be understood that the characteristic pattern hidden layer of Detail contrast network can also be replaced using other suitable modes
Calculate the characteristic pattern of original image and the characteristic pattern of the second image.It will also be appreciated that loss of correcting errors is that supervised learning obtains
A kind of comparing result obtained, loss of detail are also one kind of the difference of the characteristic pattern of original image and the characteristic pattern of the second image
Comparison result can be replaced loss and loss of detail result as a comparison of correcting errors by other values.
In one embodiment, discrimination model is made of multilayer convolutional neural networks and other multilayer deep neural networks,
By in original image and the second image (sample) input discrimination model, discrimination model output model is lost, i.e., pair of two kinds images
Compare result.Model loss be defined as judge input sample whether be original image probability, i.e. the image of generation model generation
It is judged as the cross entropy that no probability and original image are judged as the probability for being.Discrimination model is by carrying out loss function
Gradient decline and back-propagation method training whole network.In one embodiment, the characteristic pattern of original image and the second image
Characteristic pattern (feature map) can be by trained network (for example, using the residual error of ImageNet data set training
(ResNet) convolutional layer of convolutional neural networks disaggregated model) it generates, the image detail for generating model reduction can be made more true to nature.
In another embodiment, loss of detail is the loss of detail for the key area being labeled, and passes through customized loss
Function enhances model performance, can carry out additional reinforcing on provincial characteristics figure, by marking the important area of sample, adds
Thus the loss weight of the difference of the characteristic pattern of strong corresponding region reinforces generating model to the degree of concern and reduction in this kind of region
Ability increases the details damage of character area for example, being labeled to the character area of the sample of the image data base of file photocopy
Punishment is lost, keeps generation model more sensitive to character area, to protrude prior information according to demand.
Step S203: correcting errors loss and loss of detail and train and generate model and discrimination model according to acquisition.It should be understood that
The comparing result of difference comes between the other characterization original images and the second image that can also be corrected errors except loss and loss of detail
Training generates model and discrimination model.
In one embodiment, correct errors loss and the loss of detail of acquisition are lost as model, uses backpropagation side
Method generates model and discrimination model to train.
Step S204: restoring compressed image using housebroken generation model, and compressed image can be with here
It obtains, can also be obtained by other suitable methods from suitable image data base.
It should be understood that step S204 is not required, only lists user here and need with housebroken generation mould
Type is come the step that carries out when restoring the image of larger compression ratio or needing to verify the housebroken practical effect for generating model
Suddenly.
By the confrontation mode of above-mentioned generation model and discrimination model train come generation model, make it possible to restore more
The image of big compression ratio (for example, 16 times of compressions or the compression of more high power), and enable to the image detail quality restored compared with
It is good, clarity is higher, and can guarantee the readability of image, main information can be completely restored to, and then be conducive to passing through height
The image that data volume becomes smaller after compression ratio compression carries out transimission and storage management, saves bandwidth and memory space requirements.It is suitable for
Image data amount is big, requires not being very high but have certain requirements scene to information preservation to image detail.
Fig. 3 shows the schematic diagram of the image processing apparatus neural network based 300 of embodiment according to the present invention.Dress
Setting 300 may include: memory 301 and the processor 302 for being coupled to memory 301.Memory 301 for storing instruction, is located
Reason device 302 is configured as the instruction stored based on memory 301 come in the step of realization for method described in Fig. 1 and Fig. 2
One or more of any step.
As shown in figure 3, device 300 can also include communication interface 303, for carrying out information exchange with other equipment.This
Outside, device 300 can also include bus 304, and memory 301, processor 302 and communication interface 303 are by bus 304 come each other
It is communicated.
Memory 301 may include volatile memory, also may include nonvolatile memory.Processor 302 can be with
It is central processing unit (CPU), microcontroller, specific integrated circuit (ASIC), digital signal processor (DSP), field-programmable
Gate array (FPGA) or other programmable logic device or the one or more collection for being configured as realization the embodiment of the present invention
At circuit.
Alternatively, above-mentioned image processing method neural network based can be that is, tangible by computer program product
Computer readable storage medium embody.Computer program product may include computer readable storage medium, containing
For executing the computer-readable program instructions of various aspects of the disclosure.Computer readable storage medium, which can be, to be kept
The tangible device of the instruction used with storage by instruction execution equipment.Computer readable storage medium can be for example but not limited to
Storage device electric, magnetic storage apparatus, light storage device, electric magnetic storage apparatus, semiconductor memory apparatus or above-mentioned any conjunction
Suitable combination.The more specific example (non exhaustive list) of computer readable storage medium includes: portable computer diskette, hard
It is disk, random access memory (RAM), read-only memory (ROM), erasable programmable read only memory (EPROM or flash memory), quiet
State random access memory (SRAM), Portable compressed disk read-only memory (CD-ROM), digital versatile disc (DVD), memory
Stick, floppy disk, mechanical coding equipment, the punch card for being for example stored thereon with instruction or groove internal projection structure and above-mentioned times
The suitable combination of meaning.Computer readable storage medium used herein above is not interpreted instantaneous signal itself, such as radio
The electromagnetic wave of wave or other Free propagations, the electromagnetic wave propagated by waveguide or other transmission mediums are (for example, by optical fiber electricity
The light pulse of cable) or pass through electric wire transmit electric signal.
It should be noted that the above list is only specific embodiments of the present invention, it is clear that the present invention is not limited to above real
Example is applied, there are many similar variations therewith.If those skilled in the art directly exported from present disclosure or
All deformations associated, are within the scope of protection of the invention.
Claims (11)
1. a kind of image processing method neural network based, comprising:
Operation is filled to the first image using generation model to obtain the second image, wherein the first image is to original
Beginning image carries out the image generated after size compression, and the size of second image and the size of the original image are at specified ratio
Example；
The first comparing result is obtained to the original image and second image progress supervised learning using discrimination model,
And details of use contrast model calculates the characteristic pattern of the original image and the characteristic pattern of second image, and then determines
Second comparing result, wherein second comparing result indicates the characteristic pattern of the original image and the spy of second image
Levy the difference of figure；
The generation model and the discrimination model are trained based on first comparing result and second comparing result.
2. the method according to claim 1, wherein the method also includes: using training after generation model
To restore third image, wherein the third image is compressed image.
3. the method according to claim 1, wherein the ruler of the size of second image and the original image
It is very little identical.
4. the method according to claim 1, wherein being filled operation to the first image using model is generated
Obtaining the second image includes: to obtain the second image using model is generated to the first image progress multilayer convolution algorithm.
5. the method according to claim 1, wherein based on first comparing result and the second comparison knot
Fruit trains the generation model and the discrimination model includes: to make first comparing result and second comparing result
For model loss, the generation model and the discrimination model are trained using back-propagation method.
6. and described according to the method described in claim 5, it is characterized in that, first comparing result is to correct errors loss
Second comparing result is loss of detail.
7. according to the method described in claim 6, it is characterized in that, the loss of detail is for the key area being labeled
Loss of detail.
8. the method according to claim 1, wherein the spy of the Detail contrast model based on Detail contrast network
Sign figure hidden layer constructs.
9. the method according to claim 1, wherein the generation model and/or the judgment models are based on volume
Neural network is accumulated to construct.
10. a kind of image processing apparatus neural network based characterized by comprising
Memory, for storing instruction；And
Processor, is coupled to the memory, and described instruction makes described device execute root when being executed by the processor
According to method described in claim 1-9.
11. a kind of computer readable storage medium, the storage medium includes instruction, and described instruction is performed, so that described
The processor of computer executes method described in the claim 1-9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811212227.7A CN109377532B (en) | 2018-10-18 | 2018-10-18 | Image processing method and device based on neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811212227.7A CN109377532B (en) | 2018-10-18 | 2018-10-18 | Image processing method and device based on neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109377532A true CN109377532A (en) | 2019-02-22 |
| CN109377532B CN109377532B (en) | 2023-01-31 |
Family
ID=65400132
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811212227.7A Active CN109377532B (en) | 2018-10-18 | 2018-10-18 | Image processing method and device based on neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109377532B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109920016A (en) * | 2019-03-18 | 2019-06-21 | 北京市商汤科技开发有限公司 | Image generating method and device, electronic equipment and storage medium |
| CN110225350A (en) * | 2019-05-30 | 2019-09-10 | 西安电子科技大学 | Natural image compression method based on production confrontation network |
| CN112116083A (en) * | 2019-06-20 | 2020-12-22 | 地平线(上海)人工智能技术有限公司 | Neural network accelerator and detection method and device thereof |
| CN112288032A (en) * | 2020-11-18 | 2021-01-29 | 上海依图网络科技有限公司 | Method and device for quantitative model training based on generation of confrontation network |
| CN113222815A (en) * | 2021-04-26 | 2021-08-06 | 北京奇艺世纪科技有限公司 | Image adjusting method and device, electronic equipment and readable storage medium |
| CN113362403A (en) * | 2021-07-20 | 2021-09-07 | 支付宝(杭州)信息技术有限公司 | Training method and device of image processing model |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6720992B1 (en) * | 1999-08-04 | 2004-04-13 | Wolfgang Zahn | Method for improving the contrast reproduction of digitized images |
| AU2015207873A1 (en) * | 2005-11-15 | 2015-08-20 | Bernadette Garner | Method for training neural networks |
| CN107463989A (en) * | 2017-07-25 | 2017-12-12 | 福建帝视信息科技有限公司 | A kind of image based on deep learning goes compression artefacts method |
| CN108509961A (en) * | 2017-02-27 | 2018-09-07 | 北京旷视科技有限公司 | Image processing method and device |
-
2018
- 2018-10-18 CN CN201811212227.7A patent/CN109377532B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6720992B1 (en) * | 1999-08-04 | 2004-04-13 | Wolfgang Zahn | Method for improving the contrast reproduction of digitized images |
| AU2015207873A1 (en) * | 2005-11-15 | 2015-08-20 | Bernadette Garner | Method for training neural networks |
| CN108509961A (en) * | 2017-02-27 | 2018-09-07 | 北京旷视科技有限公司 | Image processing method and device |
| CN107463989A (en) * | 2017-07-25 | 2017-12-12 | 福建帝视信息科技有限公司 | A kind of image based on deep learning goes compression artefacts method |
Non-Patent Citations (1)
| Title |
|---|
| 胡长胜 等: "基于深度特征学习的图像超分辨率重建", 《自动化学报》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109920016A (en) * | 2019-03-18 | 2019-06-21 | 北京市商汤科技开发有限公司 | Image generating method and device, electronic equipment and storage medium |
| CN110225350A (en) * | 2019-05-30 | 2019-09-10 | 西安电子科技大学 | Natural image compression method based on production confrontation network |
| CN110225350B (en) * | 2019-05-30 | 2021-03-23 | 西安电子科技大学 | Natural image compression method based on generation type countermeasure network |
| CN112116083A (en) * | 2019-06-20 | 2020-12-22 | 地平线(上海)人工智能技术有限公司 | Neural network accelerator and detection method and device thereof |
| CN112116083B (en) * | 2019-06-20 | 2024-03-08 | 地平线(上海)人工智能技术有限公司 | Neural network accelerator and detection method and device thereof |
| CN112288032A (en) * | 2020-11-18 | 2021-01-29 | 上海依图网络科技有限公司 | Method and device for quantitative model training based on generation of confrontation network |
| CN112288032B (en) * | 2020-11-18 | 2022-01-14 | 上海依图网络科技有限公司 | Method and device for quantitative model training based on generation of confrontation network |
| CN113222815A (en) * | 2021-04-26 | 2021-08-06 | 北京奇艺世纪科技有限公司 | Image adjusting method and device, electronic equipment and readable storage medium |
| CN113362403A (en) * | 2021-07-20 | 2021-09-07 | 支付宝(杭州)信息技术有限公司 | Training method and device of image processing model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109377532B (en) | 2023-01-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109377532A (en) | Image processing method and device neural network based | |
| Nagaraj et al. | Competent ultra data compression by enhanced features excerption using deep learning techniques | |
| CN112883227B (en) | Video abstract generation method and device based on multi-scale time sequence characteristics | |
| CN112802076A (en) | Reflection image generation model and training method of reflection removal model | |
| CN112766099B (en) | Hyperspectral image classification method for extracting context information from local to global | |
| CN117576402B (en) | Deep learning-based multi-scale aggregation transducer remote sensing image semantic segmentation method | |
| CN114742985A (en) | Hyperspectral feature extraction method and device and storage medium | |
| CN114743080A (en) | Image processing method and device, terminal and storage medium | |
| WO2023207836A1 (en) | Image encoding method and apparatus, and image decompression method and apparatus | |
| CN116309148A (en) | Image restoration model training method, image restoration device and electronic equipment | |
| CN118032327B (en) | Equipment intelligent lubrication monitoring method and device based on artificial intelligence | |
| CN110503152B (en) | Two-way neural network training method and image processing method for target detection | |
| CN114782291A (en) | Training method and device of image generator, electronic equipment and readable storage medium | |
| CN113822790A (en) | Image processing method, device, equipment and computer readable storage medium | |
| CN115426075A (en) | Encoding transmission method of semantic communication and related equipment | |
| Zhou et al. | MSAR‐DefogNet: Lightweight cloud removal network for high resolution remote sensing images based on multi scale convolution | |
| CN117892059A (en) | Electric energy quality disturbance identification method based on multi-mode image fusion and ResNetXt-50 | |
| CN105684449B (en) | From the method and apparatus of the estimation of the lower quality version and abstract building original image of original image | |
| KR102368447B1 (en) | Compressing apparatus and method of trained deep artificial neural networks for video coding | |
| CN115426671B (en) | Method, system and equipment for training graphic neural network and predicting wireless cell faults | |
| WO2023118317A1 (en) | Method and data processing system for lossy image or video encoding, transmission and decoding | |
| CN115294405A (en) | Method, device, equipment and medium for constructing crop disease classification model | |
| CN115879507A (en) | Large-scale graph generation method based on deep confrontation learning | |
| CN115331081A (en) | Image target detection method and device | |
| CN114615505A (en) | Point cloud attribute compression method and device based on depth entropy coding and storage medium |
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 |