CN108416803A - A kind of scene depth restoration methods of the Multi-information acquisition based on deep neural network - Google Patents
A kind of scene depth restoration methods of the Multi-information acquisition based on deep neural network Download PDFInfo
- Publication number
- CN108416803A CN108416803A CN201810208334.6A CN201810208334A CN108416803A CN 108416803 A CN108416803 A CN 108416803A CN 201810208334 A CN201810208334 A CN 201810208334A CN 108416803 A CN108416803 A CN 108416803A
- Authority
- CN
- China
- Prior art keywords
- depth
- boundary
- image
- resolution
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 14
- 238000012549 training Methods 0.000 claims description 21
- 238000012360 testing method Methods 0.000 claims description 2
- 238000000605 extraction Methods 0.000 claims 1
- 238000012545 processing Methods 0.000 abstract description 3
- 230000000875 corresponding effect Effects 0.000 description 10
- 230000006870 function Effects 0.000 description 7
- 238000011084 recovery Methods 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002146 bilateral effect Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/50—Depth or shape recovery
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
- G06T3/4076—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution using the original low-resolution images to iteratively correct the high-resolution images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/13—Edge detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The present invention is a kind of scene depth restoration methods of the Multi-information acquisition based on deep neural network, belongs to image processing field.This method uses depth convolutional network predetermined depth image boundary, and using the boundary acquired carries out that interpolation is instructed to acquire high quality depth map.Using coloured image assistant images Boundary Prediction, it can preferably predict that the unconspicuous boundary on the depth image of low resolution, the depth image that coloured image auxiliary interpolation enables to meet the space structure of actual scene.Method program is simple, it is easy to accomplish.Depth information is asked to depth image piecemeal according to the boundary of prediction, calculating speed is fast, avoids the interference of the depth information of different zones, accuracy is high, and the high-resolution depth graph picture acquired is clear, and boundary is sharp.
Description
Technical field
The invention belongs to image processing field, it is related to using depth convolutional network predetermined depth image boundary, and use side
It instructs into row interpolation in the hope of high quality depth map on boundary, and in particular to a kind of field of the Multi-information acquisition based on deep neural network
Depth of field degree restoration methods.
Scene depth is particularly significant for natural scene understanding, is widely used in three-dimensional (3D) modeling, visualizes and automatic
Drive etc.；However the limitation of the complexity of actual scene and imaging sensor, the accuracy of the depth information of scene of acquisition with
And resolution ratio is all not sufficient to apply to actual scene.Such as the depth of second generation Kinect (Kinect2) acquisitions of current Microsoft
The resolution ratio of image is only 512 × 424, and the resolution ratio of corresponding coloured image is 1920 × 1080.General actual use is adopted
The depth information collected needs the depth information of acquisition promoting resolution ratio.
In general, a method for restoring high-resolution scene depth image is carried out using corresponding coloured image auxiliary
Bilateral interpolation carries out the recovery of depth image.Some existing methods are according to the corresponding property of texture of coloured image and depth image
Devise energy function (J.Yang, X.Ye, K.Li, C.Hou, and Y.Wang, " Color-guided depthrecovery
from RGBD data using an adaptive autoregressive model.”IEEETIP,vol.23,no.8,
Pp.3443-3458,2014), make the depth image both distribution of satisfaction value size and small resolutions of recovery by optimizing energy function
The image distribution of rate is identical, and meets the compatibility on texture.Certain methods structure boundary dictionary stores the depth of small resolution ratio
The texture of figure and high-resolution depth map texture correspondence (Jun Xie, R.S.Feris, and Ming Ting Sun,
“Edge-guidedsingledepthimagesuperresolution,”IEEETransactions on Image
Processing, vol.25, no.1, pp.428,2016), so that the depth image of big resolution ratio is acquired according to dictionary search
Then the interpolation of depth image is instructed on boundary.But the obtained boundary of this method is not smooth enough, and without using colored
The information of image, so picture quality is not high.
Coloured image provides high-resolution scene information, and abundant texture letter can be provided for the recovery of scene depth
Breath.Scene depth is less relative to actual scene texture information.The difficult point that depth information restores is the recovery of border texture.Base
In this, by deep learning, the prediction that can design a fusion color image information and deep image information is smoother
The network on boundary, and combine color image restoration scene depth image using the boundary predicted.
Invention content
The present invention is directed to overcome the deficiencies in the prior art, a kind of scene of the Multi-information acquisition based on deep neural network is deep
Spend restoration methods.It is observed that the texture structure of scene depth image is relatively simple compared to coloured image texture, restore accurate
Border issue is the difficult point that depth image restores.Based on this, this method devises fusion cromogram using the method for deep learning
The network on the smooth boundary of prediction of picture and depth image, and combine color image restoration scene depth using the boundary predicted
Image.
The technical scheme is that a kind of scene depth recovery side of the Multi-information acquisition based on deep neural network
Method, the method includes the following steps：
The first step prepares training data；
Training data includes high-resolution coloured image, the depth image of low resolution, high-resolution depth image
Corresponding boundary image.
Second step, builds Boundary Prediction network, and the coloured image of input includes two convolution by a convolution sum one
Residual error structure, obtain the characteristic pattern of colored branch；The depth image of input is realized using deconvolution operation and is up-sampled twice, is divided
Resolution size as color image resolution passes through 3 convolution before each deconvolution, wherein after twice convolution use it is residual
The structure of poor network is conducive to network convergence；After the characteristic pattern for merging high-resolution coloured image, prediction obtains corresponding to height
The boundary image of depth of resolution image；
Third walks, and builds loss function and training network；
Loss functionIt weighs on training data by coloured image I and low resolution depth map DlThe boundary result E of prediction and
The boundary E really extractedgtGap.
Wherein, Indicate that network infers process.Indicate square of 2 norms.Net
Network training process is to continue to optimize network parameter w in training data to makeConvergence obtains final
4th step predicts the boundary of the corresponding high-resolution scene depth image of the depth map of low resolution.According to survey
Coloured image I on examination collection and low resolution depth map Dl, the network that process has been trained obtains boundary
5th step, according to coloured image I and low resolution depth map DlIt carries out sub-regional interpolation or copy obtains high-resolution
Rate depth map Dh：
5-1) boundary of prediction is expanded, obtains smooth region and borderline region, smooth region directly carries out depth
The copy of value obtains the depth map D of smooth regionsmooth；
The depth map D of borderline region 5-2) is obtained into row interpolation to borderline regionedge：
Gaussian spatial distance d 5-2-1) is calculated between the adjacent pixel x, y on coloured image：
D=Gσ(Ix-Iy)
σ is the parameter of Gaussian function, value 0.5.Similitude is bigger between the value of d shows more greatly two pixels.IxAnd Iy
The value of color on coloured image at x and y is indicated respectively.
5-2-2) judge whether adjacent pixel x, y are in the both sides on depth image boundary, are formulated as 1 (x, y；E),
Indicate whether x, y are in the side on boundary in E, wherein 1 () for being judged as whether present case is true, set up anti-for 1
Then be 0.
5-2-3) judged according to above, carries out point-by-point interpolation.Equation is as follows：
Wherein,WithRespectively indicate high-resolution depth graph x at value of the low resolution depth map at y, K
For normalization factor.Indicate the point set on the low-resolution image around x.
The result of upper two step 5-3) is merged one piece and obtains ultimate depth result Dh=Dsmooth+Dedge。
The beneficial effects of the invention are as follows：
The present invention is based on the prioris that low resolution depth image is mainly obscured in borderline region, pass through neural network forecast height
The boundary of depth of resolution image, and then interpolation algorithm is used, high quality depth image is obtained, is had the characteristics that：
1, program is simple, it is easy to accomplish, the depth image of high quality can be obtained；
2, the recovery of scene depth image is divided into two steps by this method, passes through the high-resolution depth image of neural network forecast first
Boundary, then, in conjunction with boundary and high-resolution image into row interpolation；
3, algorithm uses deep learning neural network forecast depth image boundary, smooth clear, in conjunction with coloured image neighbor interpolation,
Obtained depth image is clear, and boundary is sharp.
Description of the drawings
Fig. 1 is implementing procedure figure, by taking the low resolution depth image block of four times of (4 ×) down-samplings as an example.
Fig. 2 is primary data.Wherein：(a) low quality depth map (b) high-resolution color figure
The case where Fig. 3 is both sides or the homonymy that adjacent pixel is in side, wherein (a) figure is x, y is in the same of boundary
Side, (b) the bright x of (c) chart, y are in the both sides on boundary.
Fig. 4 is the restoration result of two groups of data and the comparison with other methods, wherein：(a) figure is colour-depth standards
Data set, (b) Federated filter result (Yijun Li, Jia Bin Huang, Narendra Ahuja, and Ming Hsuan
Yang, " Deep joint image filtering, " in Proc.ECCV, 2016.), (c) result of the invention.
Specific implementation mode
With reference to embodiment and attached drawing to the scene depth of the Multi-information acquisition based on deep neural network of the present invention
Restoration methods are described in detail.
A kind of scene depth restoration methods of the Multi-information acquisition based on deep neural network, as shown in Figure 1, the method
(by 4 × for) include the following steps：
The first step prepares primary data；
Primary data includes low resolution depth map and the high-resolution color figure with visual angle, one of which data such as Fig. 2
It is shown.For training network, data set uses Middlebury officials data (http://vision.middlebury.edu),
In 38 colour-depth images to being used to train, 6 colour-depth images are for testing.For training data, scheme from training
As in stride be 10 pixels interception 15 × 15 depth image block.Corresponding coloured image stride is 40 pixels, and image block is big
Small is 60 × 60, ultimately forms 13860 images to being used to train.
Second step builds Boundary Prediction network, as shown in Figure 1, left-half is network structure.The coloured image of input
By one residual error structure for including two convolution of a convolution sum, the characteristic pattern of colored branch is obtained.The depth image of input
After up-sampling operation twice, resolution ratio size as color image resolution, every time by primary before up-sampling
One residual error structure for including two convolution of convolution sum.The structure of residual error network is conducive to network convergence.It merges from high-resolution
After the characteristic pattern of coloured image, two parts characteristic pattern passes through cubic convolution, can predict to obtain corresponding to high-resolution depth
The boundary image of image；
Third walks, and builds loss function and training network；
Loss functionIt weighs on training data by coloured image I and low resolution depth map DlThe boundary result E of prediction and
The boundary E really extractedgtGap.
Wherein, Indicate that network infers process.Indicate square of 2 norms.Net
Network training process is to continue to optimize network parameter w in training data to makeConvergence obtains finalWhen training, together
When input low resolution depth image and corresponding high-definition picture and high-resolution boundary image, neural network forecast is gone out
Boundary image constantly relatively and automatically updates network parameter with high-resolution boundary image, trained to achieve the purpose that.When training,
Initial learning rate is set as 0.001, and every 50 periods reduce by one times.When the loss function reduction of network tends towards stability no longer
Network training can terminate when reduction.
4th step predicts the boundary of the corresponding high-resolution scene depth image of the depth map of low resolution.According to survey
Coloured image I on examination collection and low resolution depth map Dl, infer to obtain boundary by network
5th step, according to coloured image I and low resolution depth map DlIt carries out sub-regional interpolation or copy obtains high-resolution
Rate depth map Dh：
5-1) boundary of prediction is expanded, obtains smooth region and borderline region, smooth region directly carries out depth
The copy of value obtains the depth map D of smooth regionsmooth；
The depth map D of borderline region 5-2) is obtained into row interpolation to borderline regionedge：
Gaussian spatial distance d 5-2-1) is calculated between the adjacent pixel x, y on coloured image：
D=Gσ(Ix-Iy)
σ is the parameter of Gaussian function, value 0.5.Similitude is bigger between the value of d shows more greatly two pixels.IxAnd Iy
The value of color on coloured image at x and y is indicated respectively.
5-2-2) judge whether adjacent pixel x, y are in the both sides on depth image boundary, are formulated as 1 (x, y；E),
Indicate whether x, y are in the side on boundary in E, wherein 1 () for being judged as whether present case is true, set up anti-for 1
Then be 0.It is divided into 3 kinds of situations and as shown in Figure 3：
A) line of point x, y, without overlapping or intersecting, are set up at this time with boundary；
B) line of point x, y intersect with boundary, invalid at this time；
C) line of point x, y have that pixel is overlapping with boundary, invalid at this time.
5-2-3) judged according to above, carries out point-by-point interpolation.Equation is as follows：
Wherein,WithRespectively indicate high-resolution depth graph x at value of the low resolution depth map at y, K
For normalization factor.Indicate the point set on the low-resolution image around x.
The result of upper two step 5-3) is merged one piece and obtains ultimate depth result Dh=Dsmooth+Dedge。
The restoration result of one group of data of this method pair and with the comparisons of other methods as shown in figure 4, wherein (a) figure is color
Color-depth standards data set, (b) Federated filter result (Yijun Li, Jia Bin Huang, Narendra Ahuja, and
Ming Hsuan Yang, " Deep joint image filtering, " in Proc.ECCV, 2016.), (c) of the invention
As a result.
Claims (2)
1. a kind of scene depth restoration methods of the Multi-information acquisition based on deep neural network, which is characterized in that including as follows
Step：
The first step prepares training data, including the depth image of high-resolution coloured image, low resolution and high-resolution
The corresponding boundary image of depth image；
Second step, builds Boundary Prediction network, and the coloured image of input includes the residual of two convolution by a convolution sum one
Poor structure obtains the characteristic pattern of colored branch；The depth image of input is up-sampled twice using deconvolution operation realization, resolution ratio
With size as color image resolution, 3 convolution are passed through before each deconvolution, wherein after twice convolution use residual error net
The structure of network；After the characteristic pattern for merging high-resolution coloured image, prediction obtain correspond to high-resolution depth graph as
Boundary image；
Third walks, and builds loss function and training network；
Loss functionIt weighs on training data by coloured image I and low resolution depth map DlThe boundary result E of prediction and true
The boundary E of extractiongtGap；
Wherein, Indicate that network infers process；Indicate 2 norms；Network training process
It is to continue to optimize network parameter w in training data to makeConvergence obtains final
4th step predicts the boundary of the corresponding high-resolution scene depth image of the depth map of low resolution；
According to the coloured image I and low resolution depth map D on test setl, boundary is obtained by network
5th step, according to coloured image I and low resolution depth map DlIt carries out sub-regional interpolation or copy obtains high-resolution depth
Scheme Dh。
2. a kind of scene depth restoration methods of Multi-information acquisition based on deep neural network according to claim 1：,
It is characterized in that：
5th step, according to coloured image I and low resolution depth map DlIt carries out sub-regional interpolation or copy obtains high-resolution depth
Dh, include the following steps：
5-1) boundary of prediction is expanded, obtains smooth region and borderline region, smooth region directly carries out depth value
Copy obtains the depth map D of smooth regionsmooth；
The depth map D of borderline region 5-2) is obtained into row interpolation to borderline regionedge：
Gaussian spatial distance d 5-2-1) is calculated between the adjacent pixel x, y on coloured image
D=Gσ(Ix-Iy)
σ is the parameter of Gaussian function, value 0.5；Similitude is bigger between the value of d shows more greatly two pixels；IxAnd IyRespectively
Indicate the value of color at x and y on coloured image；
5-2-2) judge whether adjacent pixel x, y are in the both sides on depth image boundary, are formulated as 1 (x, y；E), indicate
Whether x, y are in the side on boundary in E, wherein 1 () for being judged as whether present case true, set up for 1 on the contrary then
It is 0；It is divided into three kinds of situations：
A) line of point x, y, without overlapping or intersecting, are set up at this time with boundary；
B) line of point x, y intersect with boundary, invalid at this time；
C) line of point x, y have that pixel is overlapping with boundary, invalid at this time；
5-2-3) judged according to above, carries out point-by-point interpolation；Equation is as follows：
Wherein,WithHigh-resolution depth graph is indicated respectively at x and value of the low resolution depth map at y, K are to return
One changes the factor；Indicate the point set on the low-resolution image around x；
The result of upper two step 5-3) is merged one piece and obtains ultimate depth result Dh=Dsmooth+Dedge。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810208334.6A CN108416803B (en) | 2018-03-14 | 2018-03-14 | Scene depth recovery method based on multi-information fusion of deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810208334.6A CN108416803B (en) | 2018-03-14 | 2018-03-14 | Scene depth recovery method based on multi-information fusion of deep neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108416803A true CN108416803A (en) | 2018-08-17 |
| CN108416803B CN108416803B (en) | 2020-01-24 |
Family
ID=63131442
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810208334.6A Expired - Fee Related CN108416803B (en) | 2018-03-14 | 2018-03-14 | Scene depth recovery method based on multi-information fusion of deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108416803B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110111289A (en) * | 2019-04-28 | 2019-08-09 | 深圳市商汤科技有限公司 | A kind of image processing method and device |
| CN110136061A (en) * | 2019-05-10 | 2019-08-16 | 电子科技大学中山学院 | Resolution improving method and system based on depth convolution prediction and interpolation |
| CN111062981A (en) * | 2019-12-13 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Image processing method, device and storage medium |
| CN111260711A (en) * | 2020-01-10 | 2020-06-09 | 大连理工大学 | Parallax estimation method for weakly supervised trusted cost propagation |
| CN111738921A (en) * | 2020-06-15 | 2020-10-02 | 大连理工大学 | Depth super-resolution method for multi-information progressive fusion based on depth neural network |
| CN111784659A (en) * | 2020-06-29 | 2020-10-16 | 北京百度网讯科技有限公司 | Image detection method and device, electronic equipment and storage medium |
| CN113763449A (en) * | 2021-08-25 | 2021-12-07 | 北京的卢深视科技有限公司 | Depth recovery method and device, electronic equipment and storage medium |
| CN113781538A (en) * | 2021-07-27 | 2021-12-10 | 武汉中海庭数据技术有限公司 | Image depth information fusion method and system, electronic equipment and storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120007950A1 (en) * | 2010-07-09 | 2012-01-12 | Yang Jeonghyu | Method and device for converting 3d images |
| CN102722863A (en) * | 2012-04-16 | 2012-10-10 | 天津大学 | Super-resolution reconstruction method for depth map by adopting autoregressive model |
| EP2677496A1 (en) * | 2012-06-20 | 2013-12-25 | Vestel Elektronik Sanayi ve Ticaret A.S. | Method and device for determining a depth image |
| CN105741265A (en) * | 2016-01-21 | 2016-07-06 | 中国科学院深圳先进技术研究院 | Depth image processing method and depth image processing device |
| CN106447714A (en) * | 2016-09-13 | 2017-02-22 | 大连理工大学 | Scene depth recovery method based on signal decomposition |
| CN107194893A (en) * | 2017-05-22 | 2017-09-22 | 西安电子科技大学 | Depth image ultra-resolution method based on convolutional neural networks |
| CN107680140A (en) * | 2017-10-18 | 2018-02-09 | 江南大学 | A kind of depth image high-resolution reconstruction method based on Kinect cameras |
-
2018
- 2018-03-14 CN CN201810208334.6A patent/CN108416803B/en not_active Expired - Fee Related
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120007950A1 (en) * | 2010-07-09 | 2012-01-12 | Yang Jeonghyu | Method and device for converting 3d images |
| CN102722863A (en) * | 2012-04-16 | 2012-10-10 | 天津大学 | Super-resolution reconstruction method for depth map by adopting autoregressive model |
| EP2677496A1 (en) * | 2012-06-20 | 2013-12-25 | Vestel Elektronik Sanayi ve Ticaret A.S. | Method and device for determining a depth image |
| CN105741265A (en) * | 2016-01-21 | 2016-07-06 | 中国科学院深圳先进技术研究院 | Depth image processing method and depth image processing device |
| CN106447714A (en) * | 2016-09-13 | 2017-02-22 | 大连理工大学 | Scene depth recovery method based on signal decomposition |
| CN107194893A (en) * | 2017-05-22 | 2017-09-22 | 西安电子科技大学 | Depth image ultra-resolution method based on convolutional neural networks |
| CN107680140A (en) * | 2017-10-18 | 2018-02-09 | 江南大学 | A kind of depth image high-resolution reconstruction method based on Kinect cameras |
Non-Patent Citations (6)
| Title |
|---|
| JINGYU YANG 等: "Depth super-resolution via fully edge-augmented guidance", 《2017 IEEE VISUAL COMMUNICATIONS AND IMAGE PROCESSING (VCIP)》 * |
| JUN XIE 等: "Edge-Guided Single Depth Image Super Resolution", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| WENTIAN ZHOU 等: "Guided deep network for depth map super-resolution: How much can color help?", 《2017 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP)》 * |
| YAN DONG 等: "Depth map upsampling using joint edge-guided convolutional neural network for virtual view synthesizing", 《JOURNAL OF ELECTRONIC IMAGING》 * |
| YIJUN LI 等: "Deep Joint Image Filtering", 《ECCV2016》 * |
| 隋瑶 等: "基于改进的测地线距离变换的深度图像恢复", 《信息技术》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110111289A (en) * | 2019-04-28 | 2019-08-09 | 深圳市商汤科技有限公司 | A kind of image processing method and device |
| CN110111289B (en) * | 2019-04-28 | 2021-09-28 | 深圳市商汤科技有限公司 | Image processing method and device |
| CN110136061A (en) * | 2019-05-10 | 2019-08-16 | 电子科技大学中山学院 | Resolution improving method and system based on depth convolution prediction and interpolation |
| CN110136061B (en) * | 2019-05-10 | 2023-02-28 | 电子科技大学中山学院 | Resolution improving method and system based on depth convolution prediction and interpolation |
| CN111062981A (en) * | 2019-12-13 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Image processing method, device and storage medium |
| CN111062981B (en) * | 2019-12-13 | 2023-05-05 | 腾讯科技（深圳）有限公司 | Image processing method, device and storage medium |
| CN111260711B (en) * | 2020-01-10 | 2021-08-10 | 大连理工大学 | Parallax estimation method for weakly supervised trusted cost propagation |
| CN111260711A (en) * | 2020-01-10 | 2020-06-09 | 大连理工大学 | Parallax estimation method for weakly supervised trusted cost propagation |
| CN111738921A (en) * | 2020-06-15 | 2020-10-02 | 大连理工大学 | Depth super-resolution method for multi-information progressive fusion based on depth neural network |
| CN111784659A (en) * | 2020-06-29 | 2020-10-16 | 北京百度网讯科技有限公司 | Image detection method and device, electronic equipment and storage medium |
| CN113781538A (en) * | 2021-07-27 | 2021-12-10 | 武汉中海庭数据技术有限公司 | Image depth information fusion method and system, electronic equipment and storage medium |
| CN113781538B (en) * | 2021-07-27 | 2024-02-13 | 武汉中海庭数据技术有限公司 | Image depth information fusion method, system, electronic equipment and storage medium |
| CN113763449A (en) * | 2021-08-25 | 2021-12-07 | 北京的卢深视科技有限公司 | Depth recovery method and device, electronic equipment and storage medium |
| CN113763449B (en) * | 2021-08-25 | 2022-08-12 | 合肥的卢深视科技有限公司 | Depth recovery method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108416803B (en) | 2020-01-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108416803A (en) | A kind of scene depth restoration methods of the Multi-information acquisition based on deep neural network | |
| CN111062872B (en) | Image super-resolution reconstruction method and system based on edge detection | |
| CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
| CN109671023A (en) | A kind of secondary method for reconstructing of face image super-resolution | |
| CN112733950A (en) | Power equipment fault diagnosis method based on combination of image fusion and target detection | |
| CN110310319B (en) | Illumination-separated single-view human body clothing geometric detail reconstruction method and device | |
| CN108389226A (en) | A kind of unsupervised depth prediction approach based on convolutional neural networks and binocular parallax | |
| CN103020898B (en) | Sequence iris image super resolution ratio reconstruction method | |
| CN106339996B (en) | A kind of Image Blind deblurring method based on super Laplace prior | |
| CN109978786A (en) | A kind of Kinect depth map restorative procedure based on convolutional neural networks | |
| CN106067161A (en) | A kind of method that image is carried out super-resolution | |
| CN106023230B (en) | A kind of dense matching method of suitable deformation pattern | |
| CN112734890B (en) | Face replacement method and device based on three-dimensional reconstruction | |
| CN110009722A (en) | Three-dimensional rebuilding method and device | |
| CN109272447A (en) | A kind of depth map super-resolution method | |
| CN109242834A (en) | It is a kind of based on convolutional neural networks without reference stereo image quality evaluation method | |
| CN114897742B (en) | Image restoration method with texture and structural features fused twice | |
| CN111696033A (en) | Real image super-resolution model and method for learning cascaded hourglass network structure based on angular point guide | |
| CN114049464A (en) | Reconstruction method and device of three-dimensional model | |
| CN111539888A (en) | Neural network image defogging method based on pyramid channel feature attention | |
| CN103971354A (en) | Method for reconstructing low-resolution infrared image into high-resolution infrared image | |
| CN112163996A (en) | Flat-angle video fusion method based on image processing | |
| CN107958489B (en) | Curved surface reconstruction method and device | |
| CN111260706B (en) | Dense depth map calculation method based on monocular camera | |
| CN104182931B (en) | Super resolution method and device |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200124 Termination date: 20210314 |