Summary of the invention
In order to solve the above technical problems, the present invention provides a kind of 2D based on deep learning and turns 3D unit pixel block depth
The amending method and device of the depth formatted data extracted in figure, can effectively improve the accurate of 3D depth image
Property, to improve the effect that 2D turns 3D rear lens, meet the needs of spectators.
Turn to extract in 3D unit pixel block depth map in a first aspect, the present invention provides a kind of 2D based on deep learning
Depth formatted data amending method, comprising:
It imports deep learning network 2D and turns the depth formatted data extracted in the unit pixel block depth map of 3D,
Point diagram is generated, including being stored in depth formatted data in the depth point of point diagram, generates corresponding depth map, and use
The position of color point mark position depth point and state;
The first instruction for receiving user, the depth according to first instruction by smearing to the depth point in point diagram
Information is modified;
The second instruction for receiving user, according to second instruction to the position in point diagram by the way of images match tracking
The depth and location information for setting depth point are modified;
Receive user third instruction, according to the third instruction in point diagram use images match track by the way of can not
The depth point of tracking is moved;
The 4th instruction for receiving user selects the depth point in point diagram according to the 4th instruction, and will
Any non-selected depth point deletion in the convex closure of depth point composition in all life cycles of selected location depth point;
The 5th instruction for receiving user according to the position in the 5th instruction selected element figure, and adds in selected position
Depth point;
By modified point diagram with the export of depth formatted data, for making during deep learning network 2D turns 3D
It is used with the process of said units block of pixels depth map.
Optionally, the color point includes: the square point, red of the square point of yellow, the square point of green, red square point, purple
Color single arrow and green double-head arrow;
Side's point of the yellow indicates isolated positions depth point in the point diagram;
Side's point of the green indicates the part that continuous position depth point is good in the point diagram, be called it is better, it is better existing
There is depth in position again;
Red side's point indicates the part that continuous position depth point is bad in the point diagram, is called bad point, bad point only has
Position does not have depth;
The red single arrow expression respectively refers to the frame of front according to direction or subsequent frame is bad point, is in present frame
It is better;The red single arrow is also represented by the depth point for not catching up with rollback, indicates that above or below is not caught up with;
The green double-head arrow indicates the depth point newly added, and life cycle is only in present frame, i.e. isolated positions depth
Point;
Side's point of the purple indicates the depth point chosen, if in the life cycle of position depth point, display
Present frame should position show the position where nearest position key frame if having exceeded life cycle;
Wherein, position refers to that this frame itself is that position key frame or this frame are not position key frame but before it
There is position key frame below;Depth refer to this frame itself be depth key frame or this frame be not depth key frame but
It is that its front and back has depth key frame.
Optionally, it is being modified by smearing to the depth of the depth point in point diagram according to first instruction
When, further includes:
Display color point is chosen whether according to first instruction, if display grayscale image, if display original image, if aobvious
Show alternating graph, if display contour map, if be superimposed semi-transparent depth map.
Optionally, the mode of described image matched jamming, comprising: global unidirectionally isolated formula tracking, local unidirectionally isolated formula
Tracking, global unidirectionally continous way tracking, the unidirectional continous way tracking in part, the overall situation intersect isolated formula tracking and global intersection continous way
Tracking;Wherein:
The global unidirectionally isolated formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information is deleted the depth point of the frame of tracking, is saved in a manner of isolated positions depth point;
The part unidirectionally isolates formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information deletes the depth point in the Convex range for the depth point composition that tracking frame is caught up with, with isolated
The mode of depth point saves;
The global unidirectional continous way tracking, comprising: the position of the frame of tracking is deleted in unidirectional tracking, the position of assignment tracking
Depth point is set, is saved in a manner of continuous position depth point, the depth point rollback not catching up with goes to start, and marks tracking side
To red unidirectional arrow, found at the end of tracking nearest depth point depth information assign depth value;
The unidirectional continous way tracking in part, comprising: what tracking frame was caught up with deleted in unidirectional tracking, the position of assignment tracking
Depth point in the Convex range of depth point composition, is saved, the position not catching up in a manner of continuous position depth point
It sets depth point rollback to go to start, the red unidirectional arrow in label tracking direction finds nearest depth at the end of tracking
The depth of point assigns depth value;
The isolated formula tracking of global intersection, comprising: carry out twice of global unidirectional isolated formula tracking, but only deleted in first pass
Except depth point;
The global continous way of intersecting is tracked, comprising: is carried out twice of global unidirectional continous way tracking, but is only deleted in first pass
Except depth point.
Second aspect, the present invention provide a kind of 2D based on deep learning and turn to extract in 3D unit pixel block depth map
Depth formatted data modification device, comprising:
Import modul turns the position extracted in the unit pixel block depth map of 3D for importing deep learning network 2D
Depth format data generate point diagram, including being stored in depth formatted data in the depth point of point diagram, generate corresponding
Depth map, and position and state with color point mark position depth point;
Module is smeared, for receiving the first instruction of user, according to first instruction by smearing to the position in point diagram
The depth information for setting depth point is modified;
Tracking module, for receiving the second instruction of user, according to second instruction using the side of images match tracking
Formula modifies to the depth and location information of the depth point in point diagram;
Mobile module, for receive user third instruct, according to the third instruction in point diagram use images match
The depth point that the mode of tracking can not track is moved;
Removing module, for receiving the 4th instruction of user, according to the 4th instruction to the depth point in point diagram
Any non-selected position in convex closure for being selected, and the depth point in all life cycles of selected location depth point being formed
Set depth point deletion;
It adds some points module, for receiving the 5th instruction of user, according to the position in the 5th instruction selected element figure, and
Selected position adds depth point;
Export module, for exporting modified point diagram with the format of depth, in deep learning network 2D
Turn to use the process of said units block of pixels depth map to use during 3D.
Optionally, the color point includes: the square point, red of the square point of yellow, the square point of green, red square point, purple
Color single arrow and green double-head arrow;
Side's point of the yellow indicates isolated positions depth point in the point diagram;
Side's point of the green indicates the part that continuous position depth point is good in the point diagram, be called it is better, it is better existing
There is depth in position again;
Red side's point indicates the part that continuous position depth point is bad in the point diagram, is called bad point, bad point only has
Position does not have depth;
The red single arrow expression respectively refers to the frame of front according to direction or subsequent frame is bad point, is in present frame
It is better;The red single arrow is also represented by the depth point for not catching up with rollback, indicates that above or below is not caught up with;
The green double-head arrow indicates the depth point newly added, and life cycle is only in present frame, i.e. isolated positions depth
Point;
Side's point of the purple indicates the depth point chosen, if in the life cycle of position depth point, display
Present frame should position show the position where nearest position key frame if having exceeded life cycle;
Wherein, position refers to that this frame itself is that position key frame or this frame are not position key frame but before it
There is position key frame below;Depth refer to this frame itself be depth key frame or this frame be not depth key frame but
It is that its front and back has depth key frame.
Optionally, the smearing module, is also used to
Display color point is chosen whether according to first instruction, if display grayscale image, if display original image, if aobvious
Show alternating graph, if display contour map, if be superimposed semi-transparent depth map.
Optionally, the mode of described image matched jamming, comprising: global unidirectionally isolated formula tracking, local unidirectionally isolated formula
Tracking, global unidirectionally continous way tracking, the unidirectional continous way tracking in part, the overall situation intersect isolated formula tracking and global intersection continous way
Tracking;Wherein:
The global unidirectionally isolated formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information is deleted the depth point of the frame of tracking, is saved in a manner of isolated positions depth point;
The part unidirectionally isolates formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information deletes the depth point in the Convex range for the depth point composition that tracking frame is caught up with, with isolated
The mode of depth point saves;
The global unidirectional continous way tracking, comprising: the position of the frame of tracking is deleted in unidirectional tracking, the position of assignment tracking
Depth point is set, is saved in a manner of continuous position depth point, the depth point rollback not catching up with goes to start, and marks tracking side
To red unidirectional arrow, found at the end of tracking nearest depth point depth information assign depth value;
The unidirectional continous way tracking in part, comprising: what tracking frame was caught up with deleted in unidirectional tracking, the position of assignment tracking
Depth point in the Convex range of depth point composition, is saved, the position not catching up in a manner of continuous position depth point
It sets depth point rollback to go to start, the red unidirectional arrow in label tracking direction finds nearest depth at the end of tracking
The depth of point assigns depth value;
The isolated formula tracking of global intersection, comprising: carry out twice of global unidirectional isolated formula tracking, but only deleted in first pass
Except depth point;
The global continous way of intersecting is tracked, comprising: is carried out twice of global unidirectional continous way tracking, but is only deleted in first pass
Except depth point.
As shown from the above technical solution, the 2D of the invention based on deep learning turns to extract in 3D unit pixel block depth map
The amending method and device of depth formatted data out, can effectively improve the accuracy of 3D depth image, to mention
High 2D turns the effect of 3D rear lens, meets the needs of spectators.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, the technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is only
It is only a part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiment of the present invention, ordinary skill people
Member's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
Following is related names used in the embodiment of the present invention:
1, original image: need to be converted into the original image of left and right figure.
2, depth: for describing the distance of pixel distance observer.
3, it depth map: for describing the picture of image pixel depth, is indicated herein with the grayscale image of black and white.
4, a series of original image sequence frame: the original image of having time sequencings.
5, the abscissa, ordinate and depth value of corresponding original image depth point: are had recorded.It is divided into isolated positions depth
Point and continuous position depth point.It only works in a certain frame isolated positions depth point.Continuous position depth point is in continuous multiframe
It works.
6, color point: being the dot for possessing different colours and geometry, for indicating position and the shape of depth point
State.
7, it depth formatted data: is recorded with the format of abscissa, ordinate, depth value, and each frame is marked
Original image corresponding one group of abscissa, ordinate, depth value.
8, point diagram: for storing and modifying in " method and system that a kind of 2D image based on deep learning turns 3D rendering "
The depth formatted data extracted in the unit pixel depth map of generation includes the corresponding one group of depth point of each frame,
The depth point of one depth map and one group of color point, point diagram is used to record position depth format data, the depth map of point diagram
It is the depth point nearest with it to be found with each of which pixel, and assign the depth value of nearest depth point to the picture
Element is the depth information for embodying depth point.Color point: being the location information and shape for embodying depth point
State.
9, original image pixel displacement respective distance left and right figure: is used for the picture of left eye and right eye viewing according to depth map.
10, life cycle: the frame range that the depth point in point diagram works.
The 2D based on deep learning that Fig. 1 shows one embodiment of the invention offer turns to mention in 3D unit pixel block depth map
The flow diagram of the amending method for the depth formatted data taken out, as shown in Figure 1, the present embodiment based on depth
The 2D of habit turns the amending method of the depth formatted data extracted in 3D unit pixel block depth map, including step 101-
107:
101, it imports deep learning network 2D and turns the depth format extracted in the unit pixel block depth map of 3D
Data generate point diagram including being stored in depth formatted data in the depth point of point diagram and generate corresponding depth
Figure, and position and state with color point mark position depth point, can refer to Fig. 9.
In a particular application, the color point includes: the side of the square point of yellow, the square point of green, red square point, purple
Point, red single arrow and green double-head arrow;
Side's point of the yellow indicates isolated positions depth point in the point diagram;
Side's point of the green indicates the part that continuous position depth point is good in the point diagram, be called it is better, it is better existing
There is a depth in position again, position refer to this frame itself be position key frame or this frame be not position key frame but it before
There is position key frame below;Depth refer to this frame itself be depth key frame or this frame be not depth key frame but
It is that its front and back has depth key frame;
Red side's point indicates the part that continuous position depth point is bad in the point diagram, is called bad point, bad point only has
Position does not have depth;
The red single arrow expression respectively refers to the frame of front according to direction or subsequent frame is bad point, is in present frame
It is better, for playing suggesting effect;The red single arrow is also represented by the depth point for not catching up with rollback, indicate front or
It does not catch up with below;
The green double-head arrow indicates the depth point newly added, and life cycle is only in present frame, i.e. isolated positions depth
Point, for playing suggesting effect;
Side's point of the purple indicates the depth point chosen, if in the life cycle of position depth point, display
Present frame should position show the position where nearest position key frame if having exceeded life cycle, this be in order to
Can be with shift position depth point outside life cycle, for example in first frame life cycle is had selected only in the depth of first frame
Point, is switched to the second frame, and the second frame shows the position where first frame depth point.
It will be appreciated that each point manual setting life cycle is very heavy labor since the point in point diagram has very much
It is dynamic, so the range of life cycle can be indicated with the position key frame of serial number minimum and maximum.
102, the first instruction for receiving user, according to first instruction by smearing to the depth point in point diagram
Depth information is modified.
It will be appreciated that it is essential for assigning depth to if to modify depth.Since the position in point diagram is deep
There are many degree point, and each depth point setting depth is very onerous toil, so picture can be used according to first instruction
The mode of pen assigns depth value.Left button is smeared, and the depth point in paintbrush covering assigns the depth value currently selected.Right button applies
Erasing function is smeared, the depth that present frame changes the depth point before depth is restored.It can be selected according to first instruction when smearing
It selects and whether shows color point, if display grayscale image, if display original image, if (the three-dimensional eyes of band can be seen display alternating graph
Stereoscopic effect out), if display contour map, if be superimposed semi-transparent depth map.
103, the second instruction for receiving user, according to second instruction in point diagram by the way of images match tracking
Depth point depth and location information modify.
It will be appreciated that smearing the depth that can only modify single frames, when totalframes is many or wrong more, a frame frame is smeared
It is very to consume cost of labor.Cost of labor can be reduced by the way of images match tracking.For different deep errors
Form based on tracking by relevant matches, successively proposes different trace modes, the mode of described image matched jamming, packet
Include: unidirectionally isolated formula tracking, global unidirectional continous way tracking, the unidirectional continous way in part chase after for global unidirectionally isolated formula tracking, part
The isolated formula tracking of track, global intersection and global continous way of intersecting are tracked;Wherein:
The global unidirectionally isolated formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information is deleted the depth point of the frame of tracking, is saved in a manner of isolated positions depth point;It is described global single
Stablize to the tracking of isolated formula for camera lens, but interframe shakes serious or depth value mistake camera lens, Object Depth is several in camera lens
Do not change, this tracking can not transition;
The part unidirectionally isolates formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information deletes the depth point in the Convex range for the depth point composition that tracking frame is caught up with, with isolated
The mode of depth point saves;The part unidirectionally isolates formula tracking and stablizes for scene, but interframe shake is serious or deep
The camera lens of angle value mistake, Object Depth has almost no change in camera lens, this tracking can not transition;
The global unidirectional continous way tracking, comprising: the position of the frame of tracking is deleted in unidirectional tracking, the position of assignment tracking
Depth point is set, is saved in a manner of continuous position depth point, the depth point rollback not catching up with goes to start, and marks tracking side
To red unidirectional arrow, found at the end of tracking nearest depth point depth information assign depth value;The overall situation
Unidirectional continous way tracking is stablized for camera lens, but interframe shakes serious or depth value mistake camera lens, Object Depth in camera lens
It changes;
The unidirectional continous way tracking in part, comprising: what tracking frame was caught up with deleted in unidirectional tracking, the position of assignment tracking
Depth point in the Convex range of depth point composition, is saved, the position not catching up in a manner of continuous position depth point
It sets depth point rollback to go to start, the red unidirectional arrow in label tracking direction finds nearest depth at the end of tracking
The depth of point assigns depth value;The unidirectional continous way tracking in part is stablized for scene, but interframe shakes serious or depth
It is worth the camera lens of mistake, Object Depth changes in camera lens;
The isolated formula tracking of global intersection, comprising: carry out twice of global unidirectional isolated formula tracking, but only deleted in first pass
Except depth point;It is described it is global to intersect isolated formula tracking stable for camera lens or move in parallel, but interframe shake it is serious or
The camera lens of depth value mistake, Object Depth has almost no change in camera lens, this tracking can not transition;
The global continous way of intersecting is tracked, comprising: is carried out twice of global unidirectional continous way tracking, but is only deleted in first pass
Except depth point;It is described it is global to intersect continous way tracking stable for camera lens or move in parallel, but interframe shake it is serious or
The camera lens of depth value mistake, Object Depth changes in camera lens.
Relevant noun used in the mode of following described image matched jammings:
Unidirectional: tracking is from start frame to target frame;
Intersect: tracking goes to start tracking from start frame to target frame, then from target frame, is complementary to one another from both direction
Depth information;
Isolated: the life cycle of depth point is only in a frame;
Continuous: the life cycle of depth point is in multiframe, that is to say, that has the presence of that point in multiple frames, i.e.,
Continuous position depth point.Non-key frame, which is taken, finds key frame progress linear transitions;
Global: all depth points of present frame are all the depth to be tracked points;
Part: the depth point of present frame selection is the depth to be tracked point;
Convex closure: for if use is not rigorous, the point set on two-dimensional surface is given, convex closure is exactly to connect outermost point
Convex polygonal come what is constituted, it can concentrate all points comprising point.For deleting the point of the depth within the scope of it, eliminate not just
The influence of true depth point can expand by a small margin it before;
Rollback: the tracking of the depth point is deleted when the matching degree of tracking is less than preset value in continuous position depth point
The subsequent all depth points of start frame are tracked in direction.
104, receive user third instruction, according to the third instruction in point diagram use images match track by the way of
The depth point that can not be tracked is moved.
It will be appreciated that executing step 105 is to track the depth point that can not be caught up with because of some, if do not moved,
These depth points do not have correct depth value that can cause large effect to final effect, if judgement is learnt if will have
Larger shadow can be generated to final effect by having the depth catching up with point to give correct position and depth within the scope of the frame of tracking
It rings, mobile and assignment just is carried out to position depth point, that is to say, that if it is considered to influence caused by the depth point not catching up with
Less, can not have to operate on it.
105, the 4th instruction for receiving user selects the depth point in point diagram according to the 4th instruction,
And any non-selected depth point in the convex closure for forming the depth point in all life cycles of selected location depth point
It deletes.
It will be appreciated that this step 105, which avoids a frame frame, deletes the duplication of labour a little, efficiency is improved.
106, the 5th instruction for receiving user, according to the position in the 5th instruction selected element figure, and in selected position
It sets and adds depth point.
It will be appreciated that there is no position deep in some fine structures due to the limited amount of the depth point in point diagram
Spend point, if but a portion fine structure do not assign correct depth and can have large effect to overall effect, then
Propose the mode added some points solve the problems, such as it is this.
107, modified point diagram is exported with the format of depth, for during deep learning network 2D turns 3D
It is used using the process of said units block of pixels depth map.
Has an example, as shown in Fig. 2 to Fig. 8, Fig. 2 is to be converted into three-dimensional original image, and Fig. 3 is to Fig. 2
Using the depth map of the point diagram before the modification of the present embodiment the method, Fig. 4 is to use described in the present embodiment Fig. 2 in the prior art
The depth map that unit pixel block depth map before method modification is ultimately generated by subsequent deep learning network, Fig. 5 are to use
The left and right figure that the depth map of Fig. 4 generates;Fig. 6 is the depth map for using Fig. 2 the modified point diagram of the present embodiment the method, figure
7 be to pass through subsequent depth using depth formatted data derived from the modified point diagram of the present embodiment the method to Fig. 2
The depth map that learning network ultimately generates;Fig. 8 is that the left and right generated using the depth map of Fig. 7 is schemed.By Fig. 2 to Fig. 8 it is found that this reality
It applies the 2D described in example based on deep learning and turns repairing for the depth formatted data extracted in 3D unit pixel block depth map
The effect that 2D turns camera lens in 3D can be effectively improved by changing method.
The 2D based on deep learning of the present embodiment turns the depth lattice extracted in 3D unit pixel block depth map
The amending method of formula data can effectively improve the accuracy of 3D depth image, full to improve the effect that 2D turns 3D rear lens
The demand of sufficient spectators.
The 2D based on deep learning that Figure 10 shows one embodiment of the invention offer turns in 3D unit pixel block depth map
The depth formatted data extracted modification device structural schematic diagram, as shown in Figure 10, the present embodiment based on depth
The 2D of degree study turns the modification device of the depth formatted data extracted in 3D unit pixel block depth map, comprising: leads
Enter module 11, smear module 12, tracking module 13, mobile module 14, removing module 15, module of adding some points 16 and export module 17;
Wherein:
Import modul 11 turns the position extracted in the unit pixel block depth map of 3D for importing deep learning network 2D
Depth format data are set, point diagram is generated, including being stored in depth formatted data in the depth point of point diagram, generate phase
The depth map answered, and position and state with color point mark position depth point;
Module 12 is smeared, for receiving the first instruction of user, according to first instruction by smearing in point diagram
The depth information of depth point is modified;
Tracking module 13, for receiving the second instruction of user, according to second instruction using images match tracking
Mode modifies to the depth and location information of the depth point in point diagram;
Mobile module 14, for receive user third instruct, according to the third instruction in point diagram use image
The depth point that mode with tracking can not track is moved;
Removing module 15, for receiving the 4th instruction of user, according to the 4th instruction to the depth in point diagram
Point is selected, and will be any non-selected in the convex closure of the depth point composition in all life cycles of selected location depth point
Depth point deletion;
Module of adding some points 16, for receiving the 5th instruction of user, according to the position in the 5th instruction selected element figure, and
Add depth point in selected position;
Export module 17, for exporting modified point diagram with the format of depth, in deep learning network
2D turns to use the process of said units block of pixels depth map to use during 3D.
In a particular application, the color point includes: the side of the square point of yellow, the square point of green, red square point, purple
Point, red single arrow and green double-head arrow;
Side's point of the yellow indicates isolated positions depth point in the point diagram;
Side's point of the green indicates the part that continuous position depth point is good in the point diagram, be called it is better, it is better existing
There is a depth in position again, position refer to this frame itself be position key frame or this frame be not position key frame but it before
There is position key frame below;Depth refer to this frame itself be depth key frame or this frame be not depth key frame but
It is that its front and back has depth key frame;
Red side's point indicates the part that continuous position depth point is bad in the point diagram, is called bad point, bad point only has
Position does not have depth;
The red single arrow expression respectively refers to the frame of front according to direction or subsequent frame is bad point, is in present frame
It is better, for playing suggesting effect;The red single arrow is also represented by the depth point for not catching up with rollback, indicate front or
It does not catch up with below;
The green double-head arrow indicates the depth point newly added, and life cycle is only in present frame, i.e. isolated positions depth
Point, for playing suggesting effect;
Side's point of the purple indicates the depth point chosen, if in the life cycle of position depth point, display
Present frame should position show the position where nearest position key frame if having exceeded life cycle, this be in order to
Can be with shift position depth point outside life cycle, for example in first frame life cycle is had selected only in the depth of first frame
Point, is switched to the second frame, and the second frame shows the position where first frame depth point.
It will be appreciated that since the depth point in point diagram has very much, each depth point manual setting life
Period is very onerous toil, so the range of life cycle can be indicated with the position key frame of serial number minimum and maximum.
In a particular application, the smearing module 12, it may also be used for
Display color point is chosen whether according to first instruction, if display grayscale image, if display original image, if aobvious
Show alternating graph, if display contour map, if be superimposed semi-transparent depth map.
It will be appreciated that smearing the depth that can only modify single frames, when totalframes is many or wrong more, a frame frame is smeared
It is very to consume cost of labor.Cost of labor can be reduced by the way of images match tracking.For different deep errors
Form successively proposes different trace modes based on tracking by relevant matches, in a particular application, described image matching
The mode of tracking, comprising: global unidirectionally isolated formula tracking, part unidirectionally isolated formula tracking, global unidirectional continous way tracking, part
Unidirectional continous way tracking, the isolated formula tracking of global intersection and global continous way of intersecting are tracked;Wherein:
The global unidirectionally isolated formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information is deleted the depth point of the frame of tracking, is saved in a manner of isolated positions depth point;It is described global single
Stablize to the tracking of isolated formula for camera lens, but interframe shakes serious or depth value mistake camera lens, Object Depth is several in camera lens
Do not change, this tracking can not transition;
The part unidirectionally isolates formula tracking, comprising: unidirectional tracking, the depth value of assignment initial position depth point and tracking
Obtained location information deletes the depth point in the Convex range for the depth point composition that tracking frame is caught up with, with isolated
The mode of depth point saves;The part unidirectionally isolates formula tracking and stablizes for scene, but interframe shake is serious or deep
The camera lens of angle value mistake, Object Depth has almost no change in camera lens, this tracking can not transition;
The global unidirectional continous way tracking, comprising: the position of the frame of tracking is deleted in unidirectional tracking, the position of assignment tracking
Depth point is set, is saved in a manner of continuous position depth point, the depth point rollback not catching up with goes to start, and marks tracking side
To red unidirectional arrow, found at the end of tracking nearest depth point depth information assign depth value;The overall situation
Unidirectional continous way tracking is stablized for camera lens, but interframe shakes serious or depth value mistake camera lens, Object Depth in camera lens
It changes;
The unidirectional continous way tracking in part, comprising: what tracking frame was caught up with deleted in unidirectional tracking, the position of assignment tracking
Depth point in the Convex range of depth point composition, is saved, the position not catching up in a manner of continuous position depth point
It sets depth point rollback to go to start, the red unidirectional arrow in label tracking direction finds nearest depth at the end of tracking
The depth of point assigns depth value;The unidirectional continous way tracking in part is stablized for scene, but interframe shakes serious or depth
It is worth the camera lens of mistake, Object Depth changes in camera lens;
The isolated formula tracking of global intersection, comprising: carry out twice of global unidirectional isolated formula tracking, but only deleted in first pass
Except depth point;It is described it is global to intersect isolated formula tracking stable for camera lens or move in parallel, but interframe shake it is serious or
The camera lens of depth value mistake, Object Depth has almost no change in camera lens, this tracking can not transition;
The global continous way of intersecting is tracked, comprising: is carried out twice of global unidirectional continous way tracking, but is only deleted in first pass
Except depth point;It is described it is global to intersect continous way tracking stable for camera lens or move in parallel, but interframe shake it is serious or
The camera lens of depth value mistake, Object Depth changes in camera lens.
Relevant noun used in the mode of following described image matched jammings:
Unidirectional: tracking is from start frame to target frame;
Intersect: tracking goes to start tracking from start frame to target frame, then from target frame, is complementary to one another from both direction
Depth information;
Isolated: the life cycle of depth point is only in a frame;
Continuous: the life cycle of depth point is in multiframe, that is to say, that has the presence of that point in multiple frames, i.e.,
Continuous position depth point.Non-key frame, which is taken, finds key frame progress linear transitions;
Global: all depth points of present frame are all tracking points;
Part: the depth point of present frame selection is tracking point;
Convex closure: for if use is not rigorous, the point set on two-dimensional surface is given, convex closure is exactly to connect outermost point
Convex polygonal come what is constituted, it can concentrate all points comprising point.For deleting the point of the depth within the scope of it, eliminate not just
The influence of true depth point can expand by a small margin it before;
Rollback: the tracking of the depth point is deleted when the matching degree of tracking is less than preset value in continuous position depth point
The subsequent all depth points of start frame are tracked in direction.
The 2D based on deep learning of the present embodiment turns the depth lattice extracted in 3D unit pixel block depth map
The modification device of formula data is applied in processor, can effectively improve the accuracy of 3D depth image, turn 3D to improve 2D
The effect of rear lens, meets the needs of spectators.
Device described in the present embodiment can be used for executing above method embodiment, and principle is similar with technical effect, this
Place repeats no more.
It should be noted that for device embodiment, since it is basically similar to the method embodiment, so description
Fairly simple, the relevent part can refer to the partial explaination of embodiments of method.
Those of ordinary skill in the art will appreciate that: the instruction in the above various embodiments can be realized above-mentioned in any combination
This can be accomplished by hardware associated with program instructions for all or part of the steps of each method embodiment.Program above-mentioned can deposit
It is stored in a computer-readable storage medium.When being executed, execution includes the steps that above-mentioned each method embodiment to the program;And
Storage medium above-mentioned includes: the various media that can store program code such as ROM, RAM, magnetic or disk.
Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations;To the greatest extent
Pipe present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement;And these are modified or replaceed, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme.