CN104537627B - A kind of post-processing approach of depth image - Google Patents
A kind of post-processing approach of depth image Download PDFInfo
- Publication number
- CN104537627B CN104537627B CN201510009690.1A CN201510009690A CN104537627B CN 104537627 B CN104537627 B CN 104537627B CN 201510009690 A CN201510009690 A CN 201510009690A CN 104537627 B CN104537627 B CN 104537627B
- Authority
- CN
- China
- Prior art keywords
- depth image
- point
- points
- depth
- sampling
- 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.)
- Expired - Fee Related
Links
Landscapes
- Image Processing (AREA)
Abstract
The present invention relates to digital image processing field, a kind of post-processing approach of depth image is disclosed.The present invention is up-sampled by using the relatively low classical interpolation algorithm of time complexity to depth image first, then the correction of depth image progress edge line, the borderline region reparation after up-sampling are post-processed twice.Can be while raising up-sampling depth image good subjective vision effect using technical scheme, pass through edge correction, make depth image border consistent with the geometrical boundary of corresponding color image, the quality of synthesis viewpoint is effectively improved, the obscurity boundary phenomenon that up-sampling is brought is eliminated.
Description
Technical field
The present invention relates to digital image processing field, more particularly to a kind of post-processing approach of depth image.
Background technology
It is one of most basic content, scene in computer vision research that three-dimensional scenic, which is obtained, relative to the distance of video camera
Middle each point can be represented relative to the distance of video camera with depth image, i.e., the pixel value of certain point is represented pair in depth image
Answer distance of the certain point relative to video camera in scene.Depth image characterizes the third dimension information of object, therefore in three-dimensional
There is important application in the research such as reconstruction, pattern-recognition, man-machine interaction.
The technology that current NI Vision Builder for Automated Inspection obtains scene depth image can be divided into passive light method and active light method two
Major class.The general principle of passive light method is from two same scenery of (or multiple) viewing point, to obtain under different visual angles
Perceptual image, by principle of triangulation the position deviation (i.e. parallax) between image pixel is calculated to obtain the three-dimensional letter of scenery
Breath, generates depth image.But passive light method needs strict constraints and accurately corrected, and time complexity also compared with
Height, it is less in actual applications to use.Active light method refers to that vision system to scene emitted energy, then receives scene first
To the reflected energy of institute's emitted energy, by calculating, the depth information of scene is obtained.Designed at present according to such method
Kinect video camera is exactly by projecting infrared light to measurement space, so that infrared speckle is formed in body surface, by dissipating
Spot obtains the depth information of object with the demarcation of video camera distance.Depth image can promptly be generated by such a mode, and
Cost is relatively low, but the depth image obtained has the problem of resolution ratio is relatively low compared with the color image obtained simultaneously, unfavorable
In the progress of follow-up work, its application in practice is limited.So needing to handle depth image again, to improve depth
The resolution ratio of image is spent, makes it that there is the up-sampling of same size, i.e. depth image with corresponding color image.
In the prior art, the up-sampling of depth image has some classical interpolation algorithms, such as closest interpolation algorithm
(nearest), bilinear interpolation algorithm (bilinear) and bicubic interpolation algorithm (bicubic) etc..Closest interpolation algorithm
The gray value of the input pixel of the position be mappeding to from it recently is selected as interpolation result:Though this algorithm amount of calculation
It is small, but obvious mosaic phenomenon and sawtooth effect can be produced.Bilinear interpolation algorithm is then that closest interpolation algorithm is changed
Enter, it first carries out first-order linear interpolation to horizontal direction, then carries out first-order linear interpolation in vertical direction again, integrates the two
End product is just obtained through third time interpolation.The crenellated phenomena of the up-sampling result images obtained through such a method is difficult to discover,
But the edge of image can produce blooming.Bicubic interpolation algorithm is the improvement to bilinear interpolation algorithm again, and it is not only examined
The gray value for having considered direct adjoint point treats the influence of sampled point, it is also contemplated that the influence of pixel value rate of change between adjoint point, therefore institute
The to be sampled pixel value tried to achieve is more accurate.Though this algorithm overcomes the shortcoming of first two method, edge blurry phenomenon is also obtained
To improvement, but the loss in detail of original image can be caused.
It can be seen that, it is badly in need of wanting a kind of post-processing approach of depth image at present, in the case where saving time and cost, not only
The blooming at edge can be eliminated, and does not result in the loss in detail of original image.
The content of the invention
It is an object of the invention to:A kind of post-processing approach of depth image is provided, it has not only been saved processing time, and
And, by edge correction, depth image side can be made while raising up-sampling depth image good subjective vision effect
Boundary is consistent with the border of corresponding color image, effectively improves the quality of synthesis viewpoint, eliminates the module of boundary that up-sampling is brought
Paste phenomenon.
The present invention provides a kind of post-processing approach of depth image, including:
Original depth image d is up-sampled by image capturing system, the depth of depth image D' after up-sampling is obtained
The gradient value information of angle value information and color image C;
In response to the depth value information of depth image D' after the up-sampling, the initial edge of depth image after the up-sampling is obtained
Boundary E';
In response to color image C gradient value information, school is carried out to the initial boundary E' of depth image after the up-sampling
Just, make the border of its depth image consistent with the border of corresponding color image, obtain depth image boundary graph;
Repaired using the depth image boundary graph of acquisition as with reference to the border to depth image after up-sampling.
Further, the initial boundary of depth image is corrected after the described pair of up-sampling, including:
To the depth value and color image of corresponding points of the every bit P after up-sampling on depth image D' on initial boundary
C Grad is compared, if Grad is larger and depth value is smaller, and boundary point P is geometrical boundary;Otherwise intersected
Correction is examined, correct boundary point is searched in the neighbouring region of boundary point to be corrected.
Further, the crosscheck correction includes:
Judge the boundary point P left and right sides and the gradient difference value of both sides pixel up and down, be referred to as differential horizontal and
Vertical difference, if its differential horizontal is more than or equal to vertical difference, the upward position correction of water-filling square is entered to P points, otherwise,
Carry out the position correction on vertical direction;
Wherein, the position correction in horizontal direction is:
Find corresponding points P of the P points in color image CC, with PCCentered on point, left and right respectively takes the continuous pixels of r to be
Point to be investigated, all waits that investigating point constitutes region R to be investigated;If M is a certain point to be investigated, and its horizontal gradient value is more than PC
The horizontal gradient value of point, then find corresponding points M of the M points in D'dIf, MdThe horizontal gradient value of point is more than a certain threshold value thresh,
It is correct geometrical boundary point then to think M points, by the correspondence position of position correction of the P points in E' to M points;
Position correction on vertical direction is:
Find corresponding points P of the P points in color image CC, with PCCentered on point, the continuous pixels of r are respectively taken to be up and down
Point to be investigated, all waits that investigating point constitutes region R to be investigated;If M is a certain point to be investigated, and its vertical gradient value is more than PC
The vertical gradient value of point, then find corresponding points M of the M points in D'dIf, MdThe vertical gradient value of point is more than a certain threshold value thresh,
It is correct geometrical boundary point then to think M points, by the correspondence position of position correction of the P points in E' to M points.
Further, in the horizontal direction in position correction or the position correction in vertical direction, if there is multiple M points to expire
Sufficient condition, then it is check point to take the maximum point of gradient on color diagram region R to be investigated.
Further, the depth value information in response to the depth image, i.e., judge to adopt on described with the change of depth value
The initial boundary of depth image after sample.
Further, depth image is up-sampled, including:Using the relatively low classical interpolation algorithm of time complexity to depth
Degree image is up-sampled.
Further, the classical interpolation algorithm includes:Closest interpolation algorithm or bilinear interpolation algorithm or bicubic are inserted
Value-based algorithm.
Therefore, using technical solution of the present invention, due to it is determined that during boundary point, being adopted by comparing the boundary point upper
The Grad of the depth value of corresponding points after sample on depth image D' and color image C is compared, and excludes texture region side
The interference of this process of bound pair, and crosscheck correction is carried out, search correct in the neighbouring region of boundary point to be corrected
Boundary point so that the position of boundary point is more accurate, so that the edge blurry for effectively eliminating depth image after up-sampling shows
As, and ensure that its edge details.
Brief description of the drawings
Accompanying drawing described herein is used for providing a further understanding of the present invention, constitutes the part of the application, not
Inappropriate limitation of the present invention is constituted, in the accompanying drawings:
Fig. 1 is a kind of main flow schematic diagram of the post-processing approach for depth image that the embodiment of the present invention 1 is provided;
Fig. 2 is a kind of boundary line correcting process signal of the post-processing approach for depth image that the embodiment of the present invention 2 is provided
Figure.
Embodiment
Describe the present invention in detail below in conjunction with accompanying drawing and specific embodiment, herein illustrative examples of the invention
And explanation is used for explaining the present invention, but it is not as a limitation of the invention.
Embodiment 1:
A kind of main flow schematic diagram of the post-processing approach for depth image that Fig. 1 provides for the present embodiment.
Shown in Figure 1, the post-processing approach for the depth image that the present embodiment is provided mainly is walked including following flow
Suddenly.
Step 1:Original depth image d is obtained by NI Vision Builder for Automated Inspection.
Step 2:Original depth image d is up-sampled using simple difference arithmetic, and obtains depth map after up-sampling
As D' depth value information and color image C gradient value information.The simple difference arithmetic can be that time complexity is relatively low
Classical interpolation algorithm, for example:Closest interpolation algorithm or bilinear interpolation algorithm or bicubic interpolation algorithm etc..
Step 3:According to the depth value information of depth image D' after the up-sampling, rim detection is carried out to the point in D', such as
Really the depth value mutation of the point, then illustrate the point for boundary point, so as to obtain the initial boundary E' of depth image after the up-sampling.
Basis for estimation herein is:It is generally consistent in view of depth information of the same object in depth image, and different objects
Just there is the mutation of depth value in intersection, thus depth image have border sharp and the characteristics of local smoothing method.
Step 4:With reference to color image C gradient value information, the initial boundary E' of depth image after the up-sampling is entered
Row correction, makes the border of its depth image consistent with the border of corresponding color image, obtains depth image boundary graph.
Step 5:Repair, fill out using the depth image boundary graph of acquisition as with reference to the border to depth image after up-sampling
Fill obscurity boundary region.
Embodiment 2:
Fig. 2 illustrates for a kind of boundary line correcting process of post-processing approach of depth image provided in an embodiment of the present invention
Figure.
The present embodiment is a kind of preferred scheme of embodiment 1, and due to the influence of spatial domain and pixel codomain, D' has border
The blooming in region, when to detection edge, there are a variety of possibilities the position of boundary point, and many uncertainties are carried out to detection band
Factor, and the edge extracted with and its Geometry edge common point of color image for matching it is seldom, illustrate the side extracted
The position of boundary's point is very inaccurate, so obtained boundary graph E' is also required to further correction.Make its edge with and it match
The Geometry edge of color image coincide, and this will instruct this trimming process using the gradient value information of color image.
Image Edge-Detection mainly uses the gradient information of image certain point whether to judge this point as boundary point, and one
As be to think that Grad is bigger, this point for boundary point possibility it is bigger.Based on this principle, we are from color image
Gradient information come instruct correction process.But the larger point of some gradient magnitudes is not necessarily marginal point sometimes, such as to color
Image carries out rim detection, and there are two kinds of possibilities on the border extracted, and a kind of is the geometrical boundary of object, a kind of then be image line
Manage the border in region.And in depth image, even if the corresponding color image in a certain region is texture image, but as long as this area
Domain is in same depth, then this region is smooth in depth image.Its each site depth value of same object is past
Toward difference less, the only intersection in object, or foreground object and the intersection of background, just occurs the prominent of depth value
Become, so the border of detection depth image, what is obtained is usually the geometrical boundary of object.We only need to be extracted with color image
The geometrical boundary gone out come instruct correction up-sample after depth image boundary point, therefore in timing, to exclude texture region
Interference of the border to this process.Although the position of the boundary point in E' is not very accurate, each border to be corrected
Distance between point and correct boundary point can't be too big, so during correction, we are only needed in boundary point to be corrected
Correct boundary point is searched in neighbouring region and is recorded.
The process to correction does a specific description below.To each boundary point in E' boundary images, do as follows
Processing:If a certain boundary point is P in E', we are only corrected in horizontal or vertical to P points on some direction, the correction
Direction should be consistent with the depth value mutation direction that P points correspond to d borderline regions.The specific decision procedure of orientation is as follows:It is inverse
To corresponding points p of the P points on original depth-map d is found, the left and right sides of p points and the gray scale difference of both sides pixel up and down are obtained
Value, is referred to as differential horizontal and vertical difference, if its differential horizontal is more than or equal to vertical difference, then think P point needs
The position correction in horizontal direction is carried out, otherwise, the position correction on vertical direction is carried out.
If desired the position correction in horizontal direction is carried out, then finds corresponding points P of the P points in CC, with PCCentered on point,
Left and right respectively takes r continuous picture elements to be point to be investigated, and all waits that investigating point constitutes region R to be investigated.If M needs checking to be a certain
Examine a little, and its horizontal gradient value is more than PCThe horizontal gradient value of point, then M is possible to be correct boundary point, but considers
Even in same depth in color image, due to the interference of color and vein information, the larger feelings of horizontal gradient value also occur
There is no texture information interference in condition, but depth map, so we need the depth information using D', cross validation M points whether be
Correct boundary point.Find corresponding points M of the M points in D'dIf, MdThe horizontal gradient value of point is more than a certain threshold value thresh, then recognizes
It is correct geometrical boundary point for M points, by the correspondence position of position correction of the P points in E' to M points.If there is multiple M points to meet
Condition, then it is check point to take the maximum point of horizontal gradient on color diagram region to be investigated.Position correction and water on vertical direction
Square to similar.
If position of the P points in E' is (x0,y0), the position of P points is (x after correction onceg,yg)
Work as gradx(p)≥grady(p):
(xg,yg)=(xM,yM){M|M∈R,gradx(M) > gradx(PC),gradx(Md) > thresh
Work as gradx(p) < grady(p):
(xg,yg)=(xM,yM){M|M∈R,grady(M) > grady(PC),grady(Md) > thresh
Corrected repeatedly according to upper type, untill the front and rear boundary image of correction does not change, the side finally obtained
Boundary's image is E.
As seen from the above-described embodiment, using technical solution of the present invention, due to it is determined that during boundary point, by comparing the border
The Grad of the depth value and color image C of corresponding points of the point after up-sampling on depth image D' is compared, and excludes line
Interference of the zone boundary to this process is managed, and carries out crosscheck correction, is searched in the neighbouring region of boundary point to be corrected
Correct boundary point is sought so that the position of boundary point is more accurate, so as to effectively eliminate the side of depth image after up-sampling
Edge blooming, and ensure that its edge details.
Embodiments described above, does not constitute the restriction to the technical scheme protection domain.It is any in above-mentioned implementation
Modifications, equivalent substitutions and improvements made within the spirit and principle of mode etc., should be included in the protection model of the technical scheme
Within enclosing.
Claims (4)
1. a kind of post-processing approach of depth image, it is characterised in that including:
Original depth image d is up-sampled by image capturing system, the depth value of depth image D' after up-sampling is obtained
The gradient value information of information and color image C;
In response to the depth value information of depth image D' after the up-sampling, the initial boundary of depth image D' after the up-sampling is obtained
E';
In response to color image C gradient value information, the initial boundary E' of depth image D' after the up-sampling is corrected,
Make depth image D' border after the up-sampling consistent with the border of corresponding color image, obtain depth image boundary graph;
Repaired using the depth image boundary graph of acquisition as with reference to the border to depth image after up-sampling;
The initial boundary of depth image is corrected after the described pair of up-sampling, including:
To the depth value and color image C of corresponding points of the boundary point P after up-sampling on depth image D' on initial boundary E'
Grad be compared, if color image C Grad be more than the corresponding points depth value, boundary point P be geometry
Border;Otherwise crosscheck correction is carried out, correct boundary point is searched in the neighbouring regions of boundary point P;
The crosscheck correction includes:
Judge the boundary point P left and right sides and the gradient difference value of both sides pixel up and down, be referred to as differential horizontal and vertical
Difference, if its differential horizontal is more than or equal to vertical difference, the upward position correction of water-filling square is entered to P points, otherwise, is carried out
Position correction on vertical direction;
Wherein, the position correction in horizontal direction is:
Find corresponding points P of the P points in color image CC, with PCCentered on point, left and right respectively takes r continuous pixels to need checking
Examine a little, all waits that investigating point constitutes region R to be investigated;If M is a certain point to be investigated, and its horizontal gradient value is more than PCPoint
Horizontal gradient value, then find corresponding points M of the M points in D'dIf, MdThe horizontal gradient value of point is more than a certain threshold value thresh, then recognizes
It is correct geometrical boundary point for M points, by the correspondence position of position correction of the P points in E' to M points;
Position correction on vertical direction is:
Find corresponding points P of the P points in color image CC, with PCCentered on point, r continuous pixels are respectively taken up and down to need checking
Examine a little, all waits that investigating point constitutes region R to be investigated;If M is a certain point to be investigated, and its vertical gradient value is more than PCPoint
Vertical gradient value, then find corresponding points M of the M points in D'dIf, MdThe vertical gradient value of point is more than a certain threshold value thresh, then recognizes
It is correct geometrical boundary point for M points, by the correspondence position of position correction of the P points in E' to M points.
2. the post-processing approach of depth image according to claim 1, it is characterised in that
In the position correction in position correction or vertical direction in the horizontal direction, if there are multiple M points to meet condition, take
The maximum point of gradient is check point on the region R to be investigated of color image.
3. the post-processing approach of depth image according to claim 1, it is characterised in that it is described in response to the up-sampling after
Depth image D' depth value information, obtains the initial boundary of depth image after the up-sampling, including:With depth after the up-sampling
The change of image D' depth value judges the initial boundary of depth image D' after the up-sampling.
4. the post-processing approach of depth image according to claim 1, it is characterised in that
Depth image is up-sampled, including:Depth image is carried out using time complexity relatively low interpolation algorithm to adopt
Sample, the interpolation algorithm includes:Closest interpolation algorithm or bilinear interpolation algorithm or bicubic interpolation algorithm.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510009690.1A CN104537627B (en) | 2015-01-08 | 2015-01-08 | A kind of post-processing approach of depth image |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510009690.1A CN104537627B (en) | 2015-01-08 | 2015-01-08 | A kind of post-processing approach of depth image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104537627A CN104537627A (en) | 2015-04-22 |
CN104537627B true CN104537627B (en) | 2017-11-07 |
Family
ID=52853146
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510009690.1A Expired - Fee Related CN104537627B (en) | 2015-01-08 | 2015-01-08 | A kind of post-processing approach of depth image |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104537627B (en) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105517677B (en) * | 2015-05-06 | 2018-10-12 | 北京大学深圳研究生院 | The post-processing approach and device of depth map/disparity map |
CN105163129B (en) * | 2015-09-22 | 2018-01-23 | 杭州电子科技大学 | Gradient map guiding based on depth resampling 3D HEVC decoding methods |
CN107292826B (en) * | 2016-03-31 | 2021-01-22 | 富士通株式会社 | Image processing apparatus, image processing method, and image processing device |
CN109242901B (en) | 2017-07-11 | 2021-10-22 | 深圳市道通智能航空技术股份有限公司 | Image calibration method and device applied to three-dimensional camera |
CN108550167B (en) * | 2018-04-18 | 2022-05-24 | 北京航空航天大学青岛研究院 | Depth image generation method and device and electronic equipment |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6563513B1 (en) * | 2000-05-11 | 2003-05-13 | Eastman Kodak Company | Image processing method and apparatus for generating low resolution, low bit depth images |
CN103455984A (en) * | 2013-09-02 | 2013-12-18 | 清华大学深圳研究生院 | Method and device for acquiring Kinect depth image |
-
2015
- 2015-01-08 CN CN201510009690.1A patent/CN104537627B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6563513B1 (en) * | 2000-05-11 | 2003-05-13 | Eastman Kodak Company | Image processing method and apparatus for generating low resolution, low bit depth images |
CN103455984A (en) * | 2013-09-02 | 2013-12-18 | 清华大学深圳研究生院 | Method and device for acquiring Kinect depth image |
Non-Patent Citations (2)
Title |
---|
Depth map post-processing for 3D-TV;Om Prakash Gangwal等;《Consumer Electronics, 2009. ICCE "09. Digest of Technical Papers International Conference on》;20090114;第1-2页 * |
基于联合双边滤波的深度图像增强算法;刘金荣等;《计算机工程》;20140331;第40卷(第3期);第249-252和257页 * |
Also Published As
Publication number | Publication date |
---|---|
CN104537627A (en) | 2015-04-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US9214013B2 (en) | Systems and methods for correcting user identified artifacts in light field images | |
Yang et al. | Color-guided depth recovery from RGB-D data using an adaptive autoregressive model | |
CN104537627B (en) | A kind of post-processing approach of depth image | |
JP6142611B2 (en) | Method for stereo matching and system for stereo matching | |
US9684964B2 (en) | Image processing apparatus and image processing method for determining disparity | |
US9070042B2 (en) | Image processing apparatus, image processing method, and program thereof | |
EP3311361B1 (en) | Method and apparatus for determining a depth map for an image | |
US9332247B2 (en) | Image processing device, non-transitory computer readable recording medium, and image processing method | |
CN102892021B (en) | New method for synthesizing virtual viewpoint image | |
EP2887310A1 (en) | Method and apparatus for processing light-field image | |
CN104574331A (en) | Data processing method, device, computer storage medium and user terminal | |
JP2017142613A (en) | Information processing device, information processing system, information processing method and information processing program | |
WO2017096814A1 (en) | Image processing method and apparatus | |
CN106408596A (en) | Edge-based local stereo matching method | |
KR20180040846A (en) | Setting method of edge blur for edge modeling | |
DE112014006493T5 (en) | Determine a scale of three-dimensional information | |
CN109218706B (en) | Method for generating stereoscopic vision image from single image | |
KR20170025214A (en) | Method for Multi-view Depth Map Generation | |
CN109493293A (en) | A kind of image processing method and device, display equipment | |
JP6139141B2 (en) | Appearance image generation method and appearance image generation device | |
KR101733028B1 (en) | Method For Estimating Edge Displacement Againt Brightness | |
KR101841750B1 (en) | Apparatus and Method for correcting 3D contents by using matching information among images | |
JP2020201823A (en) | Image processing device, image processing method, and program | |
JP2006023133A (en) | Instrument and method for measuring three-dimensional shape | |
JP6351364B2 (en) | Information processing apparatus, information processing method, and program |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
C10 | Entry into 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 |
Granted publication date: 20171107 Termination date: 20220108 |
|
CF01 | Termination of patent right due to non-payment of annual fee |