CN109242788A - One kind being based on coding-decoding convolutional neural networks low-light (level) image optimization method - Google Patents
One kind being based on coding-decoding convolutional neural networks low-light (level) image optimization method Download PDFInfo
- Publication number
- CN109242788A CN109242788A CN201810952196.2A CN201810952196A CN109242788A CN 109242788 A CN109242788 A CN 109242788A CN 201810952196 A CN201810952196 A CN 201810952196A CN 109242788 A CN109242788 A CN 109242788A
- Authority
- CN
- China
- Prior art keywords
- image
- light
- level
- low
- neural network
- 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
- 238000000034 method Methods 0.000 title claims abstract description 21
- 238000005457 optimization Methods 0.000 title claims abstract description 20
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 17
- 238000012549 training Methods 0.000 claims abstract description 29
- 238000003062 neural network model Methods 0.000 claims abstract description 27
- 238000005286 illumination Methods 0.000 claims abstract description 10
- 238000003384 imaging method Methods 0.000 claims description 16
- 230000006870 function Effects 0.000 claims description 10
- 238000005520 cutting process Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 description 9
- 238000012545 processing Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20024—Filtering details
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to one kind to be based on coding-decoding convolutional neural networks low-light (level) image optimization method, uses low-light (level) image with corresponding normal illumination image as training dataset first；Then by way of data-driven, using the training dataset training U-Net type neural network model of step S1, its autonomous learning data characteristics is enabled；Optimization is reconstructed to collected low-light (level) image finally by the U-Net type neural network model after training, realizes image reconstruction.The present invention can be in the case where can not accurately model restructing algorithm, important feature in autonomous learning low-light (level) image, be not necessarily to manual intervention.
Description
Technical field
The present invention relates to image procossing domains, especially a kind of excellent based on coding-decoding convolutional neural networks low-light (level) image
Change method.
Background technique
In real life, due to the limitation of environmental factor and imaging system condition, often there is many in image
Noise, these picture noises have very big interference to the accuracy that human vision analyzes and determines.Currently, many noise images are all
It can be restored well by restructing algorithm, however, under low-light level environment, since photon numbers and noise are relatively low,
It therefore under image quality is under normal conditions bad, and it is very big for optimizing having always for picture quality under low-light level environment
Market application scenarios, but up to the present under low-light level environment optimize picture quality still there is very big challenge.
Aiming at the problem that optimizing picture quality under low-light level environment, many solutions have been proposed at present.A kind of side
Case is to obtain more visible image by adjusting image system hardware, such as sensitivity is turned up to increase brightness, passes through opening
Aperture, prolonging exposure time increase in the way of flash lamp etc. signal-to-noise ratio under low-light level environment.But these methods have them
Respective disadvantage, sensitivity, which is turned up, can make noise be simultaneously amplified, and prolonging exposure time can be because camera shake or target be moved
It is dynamic to fog, therefore image quality is often bad.
Another scheme is that low-light (level) image is reconstructed by image processing algorithm, improves the brightness and noise of image
Than.A kind of effective image processing algorithm is three-dimensional block matching algorithm, which divides the image into the block of same size, root first
According to the similitude between them, it is combined into one group of three-dimensional matrice, is then handled with the method for associated filters, finally by
Inverse transformation, treated, result is returned in original image, thus the reconstructed image after being optimized, however due to this method
Belong to non-blind property reconstruct, needs to carry out manual intervention, the noise level parameters in clearly specified present image, if the noise of setting
Rank is too small, and many noises may be still remained on image, if the level of noise of setting is excessive, may cause smooth, use
Scene is very restricted.Another algorithm is the burst imaging method proposed by Hasinoff et al., and this method passes through school
Good effect is just obtained with the mode for mixing multiple low-light (level) images, but due to needing intensive compliance evaluation, centainly
The complexity of model is increased in degree, therefore is not suitable for the processing of video image.
Summary of the invention
In view of this, the purpose of the present invention is to propose to one kind based on coding-decoding convolutional neural networks low-light (level) image it is excellent
Change method, being capable of important spy in the case where can not accurately model to restructing algorithm, in autonomous learning low-light (level) image
Sign is not necessarily to manual intervention.
The present invention is realized using following scheme: one kind being based on coding-decoding convolutional neural networks low-light (level) image optimization side
Method, specifically includes the following steps:
Step S1: using low-light (level) image with corresponding normal illumination image as training dataset；
Step S2: by way of data-driven, using the training dataset training U-Net type neural network mould of step S1
Type enables its autonomous learning data characteristics；
Step S3: collected low-light (level) image is reconstructed by the U-Net type neural network model after training excellent
Change, realizes image reconstruction；Wherein, the reconstruction and optimization formula of the U-Net type neural network model after the training are as follows:
Wherein, RlearnFor by the U-Net type neural network model after study, f () is loss function, g () is regularization
Parametric function (purpose is in order to avoid over-fitting), Θ are all parameters in deep neural network model, and θ is the member in Θ
Element, xnFor the imaging under the conditions of normal light source, ynFor the imaging under low-light level environment, RθFor U-Net type neural network model, N is
Image pixel point quantity.By multiple cycle trainings, model is made to reach optimum efficiency, once training is completed, RlearnIt can pass through
Low-light (level) image optimized after reconstructed image.
Further, the U-Net type neural network model is constricted path in left-half, to extract the spy of data
Sign, right half part is path expander, to increase the dimension of characteristic pattern；In order to be accurately positioned detail textures feature, road will be shunk
The characteristic pattern of diameter is combined with the characteristic pattern of identical dimensional on symmetrical expansion path, and final output is counted up to input dimension, channel
Exactly the same data realize image reconstruction end to end.
Further, the U-Net type neural network model is full convolutional network structure, and input picture first passes around 7 layers
Convolutional layer is obtained wherein 3 × 3 convolution kernels that each convolutional layer is 1 by step-length carry out convolution twice, is carried out using padding
Zero padding operation guarantees that the input dimension of convolution is identical as dimension is exported, the characteristic pattern after each convolution using activation primitive into
Row Nonlinear Mapping；Wherein, convolutional layer 1 is known as constricted path to convolutional layer 4, use between different convolutional layers step-length for 22 ×
2 filters carry out maximum pondization operation, to extract main feature, reduce the dimension of characteristic pattern；Convolutional layer 4 claims to convolutional layer 7
For path expander, uses step-length to carry out deconvolution for 22 × 2 filters between different convolutional layers, increases the dimension of characteristic pattern,
Subsequent convolution operation is executed again after merging simultaneously with the characteristic pattern in constricted path with identical dimensional convolutional layer, is passing through 7
After layer convolutional layer, characteristic pattern identical with input dimension, 1 × 1 convolution nuclear convolution for being 1 using one layer of step-length, reconstruct are obtained
Obtain output image.
Further, step S2 specifically:
Step S21: using low-light (level) image as input, by the forward-propagating of U-Net type neural network model, one is obtained
Open output image；
Step S22: the penalty values of output image and corresponding normal illumination image are found out using L1 loss function, and are used
Adam optimizer updates model parameter value；
Step S23: entire model 3000 periods of iteration, cycle iterations terminate, then model training flexure.
Further, step S21 further include: first by the sample of low-light (level) image and normal illumination image in same zone
Domain is cut, and the image size after cutting is 256 × 256, and carries out the operation including overturning to the image after cutting
Increase data set size, prevents over-fitting.
Further, in step S23, in 3000 periods, wherein the learning rate in preceding 1500 periods is 0.0001,
The learning rate in 1500 periods is set as 0.00001 afterwards.
The present invention by the way of data-driven, using encode end to end-decode convolutional neural networks, by convolution
Neural network is trained, and obtains model parameter, can be can not be accurately to restructing algorithm compared to traditional images restructing algorithm
In the case where being modeled, important feature in autonomous learning low-light (level) image is not necessarily to manual intervention.Once convolutional neural networks
Model training is completed, and just can quickly realize low-light (level) image reconstruction.Simultaneously in the quality of optimization low-light (level) image, therewith
Preceding algorithm is compared, and has apparent advantage.
Compared with prior art, the invention has the following beneficial effects:
1, the present invention is handled image using deep neural network algorithm, can be suitable for currently on the market any
A imaging system can reduce imaging system cost compared to physical hardware optimization method.
2. the present invention optimizes the algorithm of picture quality under low-light level environment compared to tradition, using the mould of deep neural network
Type carries out image procossing, can main feature in autonomous learning data, be not necessarily to manual intervention, for optimization low-light-level imaging this
Class can not accurately construct the scene of reconstruction model, have a clear superiority.
3, the present invention has certain generalization ability, and the image for using the imaging system different from data set to obtain is as defeated
Enter, good effect can be obtained to a certain extent.
4, the present invention is suitable for the imaging operation under low-light level environment, optimizes the image matter of imaging system under low-light level environment
Amount, under low brightness condition target detection and tracking, object identification all have a good application prospect.
Detailed description of the invention
Fig. 1 is the neural network structure schematic diagram of the embodiment of the present invention.
Fig. 2 is the neural metwork training flow diagram of the embodiment of the present invention.
Fig. 3 is the using process diagram of the embodiment of the present invention.
Specific embodiment
The present invention will be further described with reference to the accompanying drawings and embodiments.
It is noted that described further below be all exemplary, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular
Also it is intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or their combination.
It present embodiments provides a kind of based on coding-decoding convolutional neural networks low-light (level) image optimization method, specific packet
Include following steps:
Step S1: using low-light (level) image with corresponding normal illumination image as training dataset；
Step S2: by way of data-driven, using the training dataset training U-Net type neural network mould of step S1
Type enables its autonomous learning data characteristics；
Step S3: collected low-light (level) image is reconstructed by the U-Net type neural network model after training excellent
Change, realizes image reconstruction；Wherein, the reconstruction and optimization formula of the U-Net type neural network model after the training are as follows:
Wherein, RlearnFor by the U-Net type neural network model after study, f () is loss function, g () is regularization
Parametric function (purpose is in order to avoid over-fitting), Θ are all parameters in deep neural network model, and θ is the member in Θ
Element, xnFor the imaging under the conditions of normal light source, ynFor the imaging under low-light level environment, RθFor U-Net type neural network model, N is
Image pixel point quantity.By multiple cycle trainings, model is made to reach optimum efficiency, once training is completed, RlearnIt can pass through
Low-light (level) image optimized after reconstructed image.
As shown in Figure 1, the model of the present embodiment be a kind of deformation of coding-decoding convolutional neural networks, entire model from
U-shaped, therefore referred to as U-Net model is showed in structure.In the present embodiment, the U-Net type neural network model is on a left side
Half portion is divided into constricted path, and to extract the feature of data, right half part is path expander, to increase the dimension of characteristic pattern；
In order to be accurately positioned detail textures feature, by the characteristic pattern phase of the characteristic pattern of constricted path and identical dimensional on symmetrical expansion path
In conjunction with final output and input dimension, the identical data of port number realize image reconstruction end to end.
In the present embodiment, the U-Net type neural network model is full convolutional network structure, and input picture first passes around
7 layers of convolutional layer, wherein each convolutional layer by step-length be 13 × 3 convolution kernels carry out twice convolution obtain, using padding into
Row zero padding operation guarantees that the input dimension of convolution is identical as output dimension, and the characteristic pattern after each convolution is using activation primitive
Carry out Nonlinear Mapping；Wherein, convolutional layer 1 is known as constricted path to convolutional layer 4, use between different convolutional layers step-length for 22
× 2 filters carry out maximum pondization operation, to extract main feature, reduce the dimension of characteristic pattern；Convolutional layer 4 arrives convolutional layer 7
Referred to as path expander uses between different convolutional layers step-length to carry out deconvolution for 22 × 2 filters, increases the dimension of characteristic pattern
Degree, while subsequent convolution operation is executed again after merging with the characteristic pattern in constricted path with identical dimensional convolutional layer, it is passing through
After crossing 7 layers of convolutional layer, characteristic pattern identical with input dimension, 1 × 1 convolution nuclear convolution for being 1 using one layer of step-length, weight are obtained
Structure obtains output image.
Preferably, the activation primitive is line rectification function.
As shown in Fig. 2, in the present embodiment, step S2 specifically:
Step S21: using low-light (level) image as input, by the forward-propagating of U-Net type neural network model, one is obtained
Open output image；
Step S22: the penalty values of output image and corresponding normal illumination image are found out using L1 loss function, and are used
Adam optimizer updates model parameter value；
Step S23: entire model 3000 periods of iteration, cycle iterations terminate, then model training flexure.
In the present embodiment, step S21 further include: first by the sample of low-light (level) image and normal illumination image in phase
It is cut with region, the image size after cutting is 256 × 256, and carries out including overturning to the image after cutting
Operation prevents over-fitting to increase data set size.
In the present embodiment, in step S23, in 3000 periods, wherein the learning rate in preceding 1500 periods is
0.0001, the learning rate in rear 1500 periods is set as 0.00001.
Particularly, it as shown in figure 3, the present embodiment acquires image by imaging system first under low-brightness scene, will adopt
The image data collected is transferred to deep neural network image processing module, and trained U-Net is contained in the module
Model parameter, reconstructed image of the input picture after deep neural network image processing module is optimized, with low-light level field
Imaging under scape is compared, and is all improved significantly from human vision and picture appraisal quality index.
The present embodiment by the way of data-driven, using encode end to end-decode convolutional neural networks, by volume
Product neural network is trained, and obtains model parameter, compared to traditional images restructing algorithm, can accurately can not calculated reconstruct
In the case that method is modeled, important feature in autonomous learning low-light (level) image is not necessarily to manual intervention.Once convolutional Neural net
Network model training is completed, and just can quickly realize low-light (level) image reconstruction.Simultaneously in the quality of optimization low-light (level) image, with
Algorithm before is compared, and has apparent advantage.
The foregoing is merely presently preferred embodiments of the present invention, all equivalent changes done according to scope of the present invention patent with
Modification, is all covered by the present invention.
Claims (6)
1. one kind is based on coding-decoding convolutional neural networks low-light (level) image optimization method, it is characterised in that: including following step
It is rapid:
Step S1: using low-light (level) image with corresponding normal illumination image as training dataset；
Step S2: by way of data-driven, training U-Net type neural network model using the training dataset of step S1,
Enable its autonomous learning data characteristics；
Step S3: being reconstructed optimization to collected low-light (level) image by the U-Net type neural network model after training, real
Existing image reconstruction；Wherein, the reconstruction and optimization formula of the U-Net type neural network model after the training are as follows:
Wherein, RlearnFor by the U-Net type neural network model after study, f () is loss function, g () is regularization parameter
Function, Θ are all parameters in deep neural network model, and θ is the element in Θ, xnFor the imaging under the conditions of normal light source,
ynFor the imaging under low-light level environment, RθFor U-Net type neural network model, N is image pixel point quantity.
2. it is according to claim 1 a kind of based on coding-decoding convolutional neural networks low-light (level) image optimization method, it is special
Sign is: the U-Net type neural network model is constricted path in left-half, to extract the feature of data, right side
It is divided into path expander, to increase the dimension of characteristic pattern；In order to be accurately positioned detail textures feature, by the characteristic pattern of constricted path
It is combined with the characteristic pattern of identical dimensional on symmetrical expansion path, final output and input dimension, the identical number of port number
According to realizing image reconstruction end to end.
3. it is according to claim 2 a kind of based on coding-decoding convolutional neural networks low-light (level) image optimization method, it is special
Sign is: the U-Net type neural network model is full convolutional network structure, and input picture first passes around 7 layers of convolutional layer, wherein
3 × 3 convolution kernels that each convolutional layer is 1 by step-length carry out convolution twice and obtain, and carry out zero padding operation using padding, protect
The input dimension for demonstrate,proving convolution is identical as output dimension, and the characteristic pattern after each convolution carries out non-linear reflect using activation primitive
It penetrates；Wherein, convolutional layer 1 is known as constricted path to convolutional layer 4, use between different convolutional layers step-length for 22 × 2 filters into
Row maximum pondization operation, to extract main feature, reduces the dimension of characteristic pattern；Convolutional layer 4 is known as expanding road to convolutional layer 7
Diameter uses between different convolutional layers step-length to carry out deconvolution for 22 × 2 filters, increases the dimension of characteristic pattern, at the same with receipts
Characteristic pattern on contracting path with identical dimensional convolutional layer executes subsequent convolution operation after merging again, is passing through 7 layers of convolutional layer
Afterwards, characteristic pattern identical with input dimension, 1 × 1 convolution nuclear convolution for being 1 using one layer of step-length are obtained, reconstruct is exported
Image.
4. it is according to claim 1 a kind of based on coding-decoding convolutional neural networks low-light (level) image optimization method, it is special
Sign is: step S2 specifically:
Step S21: using low-light (level) image as input, by the forward-propagating of U-Net type neural network model, obtain one it is defeated
Image out；
Step S22: the penalty values of output image and corresponding normal illumination image are found out using L1 loss function, and use Adam
Optimizer updates model parameter value；
Step S23: entire model 3000 periods of iteration, cycle iterations terminate, then model training is completed.
5. it is according to claim 4 a kind of based on coding-decoding convolutional neural networks low-light (level) image optimization method, it is special
Sign is: step S21 further include: first cuts out the sample of low-light (level) image and normal illumination image in same area
It cuts, the image size after cutting is 256 × 256, and carries out the operation including overturning to the image after cutting to increase number
According to collection size, over-fitting is prevented.
6. it is according to claim 4 a kind of based on coding-decoding convolutional neural networks low-light (level) image optimization method, it is special
Sign is: in step S23, in 3000 periods, wherein the learning rate in preceding 1500 periods is 0.0001, rear 1500 week
The learning rate of phase is set as 0.00001.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810952196.2A CN109242788A (en) | 2018-08-21 | 2018-08-21 | One kind being based on coding-decoding convolutional neural networks low-light (level) image optimization method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810952196.2A CN109242788A (en) | 2018-08-21 | 2018-08-21 | One kind being based on coding-decoding convolutional neural networks low-light (level) image optimization method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109242788A true CN109242788A (en) | 2019-01-18 |
Family
ID=65071688
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810952196.2A Pending CN109242788A (en) | 2018-08-21 | 2018-08-21 | One kind being based on coding-decoding convolutional neural networks low-light (level) image optimization method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109242788A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109978962A (en) * | 2019-04-09 | 2019-07-05 | 广州市交通高级技工学校（广州市交通技师学院） | A kind of low contrast indicating value digital image recognition method towards the calibrating of darkroom illumination photometer |
| CN110033419A (en) * | 2019-04-17 | 2019-07-19 | 山东超越数控电子股份有限公司 | A kind of processing method being adapted to warship basic image defogging |
| CN110060251A (en) * | 2019-04-26 | 2019-07-26 | 福州大学 | A kind of building surface crack detecting method based on U-Net |
| CN110070498A (en) * | 2019-03-12 | 2019-07-30 | 浙江工业大学 | A kind of image enchancing method based on convolution self-encoding encoder |
| CN110097106A (en) * | 2019-04-22 | 2019-08-06 | 苏州千视通视觉科技股份有限公司 | The low-light-level imaging algorithm and device of U-net network based on deep learning |
| CN110097515A (en) * | 2019-04-22 | 2019-08-06 | 苏州千视通视觉科技股份有限公司 | Low-light (level) image processing algorithm and device based on deep learning and spatio-temporal filtering |
| CN110163815A (en) * | 2019-04-22 | 2019-08-23 | 桂林电子科技大学 | Low-light (level) restoring method based on multistage variation self-encoding encoder |
| CN110378845A (en) * | 2019-06-17 | 2019-10-25 | 杭州电子科技大学 | A kind of image repair method under extreme condition based on convolutional neural networks |
| CN110706173A (en) * | 2019-09-27 | 2020-01-17 | 中国计量大学 | Atomic force microscope image blind restoration method based on convolutional neural network |
| CN110728643A (en) * | 2019-10-18 | 2020-01-24 | 上海海事大学 | Low-illumination band noise image optimization method based on convolutional neural network |
| CN111047532A (en) * | 2019-12-06 | 2020-04-21 | 广东启迪图卫科技股份有限公司 | Low-illumination video enhancement method based on 3D convolutional neural network |
| CN111931857A (en) * | 2020-08-14 | 2020-11-13 | 桂林电子科技大学 | MSCFF-based low-illumination target detection method |
| WO2020238123A1 (en) * | 2019-05-31 | 2020-12-03 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for improving color quality of images |
| US11057634B2 (en) | 2019-05-15 | 2021-07-06 | Disney Enterprises, Inc. | Content adaptive optimization for neural data compression |
| WO2021232195A1 (en) * | 2020-05-18 | 2021-11-25 | Shanghai United Imaging Healthcare Co., Ltd. | Systems and methods for image optimization |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105519109A (en) * | 2013-08-06 | 2016-04-20 | 微软技术许可有限责任公司 | Encoding video captured in low light |
| CN105844238A (en) * | 2016-03-23 | 2016-08-10 | 乐视云计算有限公司 | Method and system for discriminating videos |
| CN105894046A (en) * | 2016-06-16 | 2016-08-24 | 北京市商汤科技开发有限公司 | Convolutional neural network training and image processing method and system and computer equipment |
-
2018
- 2018-08-21 CN CN201810952196.2A patent/CN109242788A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105519109A (en) * | 2013-08-06 | 2016-04-20 | 微软技术许可有限责任公司 | Encoding video captured in low light |
| CN105844238A (en) * | 2016-03-23 | 2016-08-10 | 乐视云计算有限公司 | Method and system for discriminating videos |
| CN105894046A (en) * | 2016-06-16 | 2016-08-24 | 北京市商汤科技开发有限公司 | Convolutional neural network training and image processing method and system and computer equipment |
Non-Patent Citations (4)
| Title |
|---|
| KIN GWN LORE ET AL.: "LLNet: A Deep Autoencoder Approach to Natural Low-light Image Enhancement", 《PATTERN RECOGNITION》 * |
| OLAF RONNEBERGER ET AL.: "U-Net: Convolutional Networks for Biomedical Image Segmentation", 《ARXIV》 * |
| ZHANG, C. ET AL.: "CT artifact reduction via U-net CNN", 《SPIE 10574,MEDICAL IMAGING 2018: IMAGE PROCESSING》 * |
| 刘超等: "超低照度下微光图像的深度卷积自编码网络复原", 《光学精密工程》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110070498A (en) * | 2019-03-12 | 2019-07-30 | 浙江工业大学 | A kind of image enchancing method based on convolution self-encoding encoder |
| CN109978962A (en) * | 2019-04-09 | 2019-07-05 | 广州市交通高级技工学校（广州市交通技师学院） | A kind of low contrast indicating value digital image recognition method towards the calibrating of darkroom illumination photometer |
| CN109978962B (en) * | 2019-04-09 | 2022-05-17 | 广州市交通高级技工学校（广州市交通技师学院） | Low-contrast indicating value image intelligent identification method for darkroom illuminometer calibration |
| CN110033419A (en) * | 2019-04-17 | 2019-07-19 | 山东超越数控电子股份有限公司 | A kind of processing method being adapted to warship basic image defogging |
| CN110097106A (en) * | 2019-04-22 | 2019-08-06 | 苏州千视通视觉科技股份有限公司 | The low-light-level imaging algorithm and device of U-net network based on deep learning |
| CN110097515A (en) * | 2019-04-22 | 2019-08-06 | 苏州千视通视觉科技股份有限公司 | Low-light (level) image processing algorithm and device based on deep learning and spatio-temporal filtering |
| CN110163815A (en) * | 2019-04-22 | 2019-08-23 | 桂林电子科技大学 | Low-light (level) restoring method based on multistage variation self-encoding encoder |
| CN110163815B (en) * | 2019-04-22 | 2022-06-24 | 桂林电子科技大学 | Low-illumination reduction method based on multi-stage variational self-encoder |
| CN110060251A (en) * | 2019-04-26 | 2019-07-26 | 福州大学 | A kind of building surface crack detecting method based on U-Net |
| US11057634B2 (en) | 2019-05-15 | 2021-07-06 | Disney Enterprises, Inc. | Content adaptive optimization for neural data compression |
| WO2020238123A1 (en) * | 2019-05-31 | 2020-12-03 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for improving color quality of images |
| US11991487B2 (en) | 2019-05-31 | 2024-05-21 | Guangdong Oppo Mobile Telecommunications Corp., Ltd. | Method, system, and computer-readable medium for improving color quality of images |
| CN110378845B (en) * | 2019-06-17 | 2021-05-25 | 杭州电子科技大学 | Image restoration method based on convolutional neural network under extreme conditions |
| CN110378845A (en) * | 2019-06-17 | 2019-10-25 | 杭州电子科技大学 | A kind of image repair method under extreme condition based on convolutional neural networks |
| CN110706173A (en) * | 2019-09-27 | 2020-01-17 | 中国计量大学 | Atomic force microscope image blind restoration method based on convolutional neural network |
| CN110728643A (en) * | 2019-10-18 | 2020-01-24 | 上海海事大学 | Low-illumination band noise image optimization method based on convolutional neural network |
| CN111047532A (en) * | 2019-12-06 | 2020-04-21 | 广东启迪图卫科技股份有限公司 | Low-illumination video enhancement method based on 3D convolutional neural network |
| WO2021232195A1 (en) * | 2020-05-18 | 2021-11-25 | Shanghai United Imaging Healthcare Co., Ltd. | Systems and methods for image optimization |
| CN111931857B (en) * | 2020-08-14 | 2022-09-02 | 桂林电子科技大学 | MSCFF-based low-illumination target detection method |
| CN111931857A (en) * | 2020-08-14 | 2020-11-13 | 桂林电子科技大学 | MSCFF-based low-illumination target detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109242788A (en) | One kind being based on coding-decoding convolutional neural networks low-light (level) image optimization method | |
| US20200151858A1 (en) | Image contrast enhancement method and device, and storage medium | |
| CN108550115B (en) | Image super-resolution reconstruction method | |
| Zhou et al. | Lednet: Joint low-light enhancement and deblurring in the dark | |
| CN110378845B (en) | Image restoration method based on convolutional neural network under extreme conditions | |
| CN109712083A (en) | A kind of single image to the fog method based on convolutional neural networks | |
| CN109255758B (en) | Image enhancement method based on all 1 x 1 convolution neural network | |
| CN111915525B (en) | Low-illumination image enhancement method capable of generating countermeasure network based on improved depth separation | |
| CN107203985B (en) | A kind of more exposure image fusion methods under end-to-end deep learning frame | |
| CN112001863B (en) | Underexposure image recovery method based on deep learning | |
| CN108280814A (en) | Light field image angle super-resolution rate method for reconstructing based on perception loss | |
| CN113168670A (en) | Bright spot removal using neural networks | |
| CN111402145B (en) | Self-supervision low-illumination image enhancement method based on deep learning | |
| CN112308803B (en) | Self-supervision low-illumination image enhancement and denoising method based on deep learning | |
| Pang et al. | Fan: Frequency aggregation network for real image super-resolution | |
| CN114897752B (en) | Single-lens large-depth-of-field computing imaging system and method based on deep learning | |
| CN109584188A (en) | A kind of image defogging method based on convolutional neural networks | |
| CN106296618A (en) | A kind of color image defogging method based on Gaussian function weighted histogram regulation | |
| CN114219722A (en) | Low-illumination image enhancement method by utilizing time-frequency domain hierarchical processing | |
| CN112465726A (en) | Low-illumination adjustable brightness enhancement method based on reference brightness index guidance | |
| CN111553856B (en) | Image defogging method based on depth estimation assistance | |
| Panetta et al. | Deep perceptual image enhancement network for exposure restoration | |
| Zhang et al. | INFWIDE: Image and feature space Wiener deconvolution network for non-blind image deblurring in low-light conditions | |
| Chen et al. | Single-image hdr reconstruction with task-specific network based on channel adaptive RDN | |
| CN113112439A (en) | Image fusion method, training method, device and equipment of image fusion model |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190118 |