CN107563963A - A kind of method based on individual depth map super-resolution rebuilding - Google Patents
A kind of method based on individual depth map super-resolution rebuilding Download PDFInfo
- Publication number
- CN107563963A CN107563963A CN201710686263.6A CN201710686263A CN107563963A CN 107563963 A CN107563963 A CN 107563963A CN 201710686263 A CN201710686263 A CN 201710686263A CN 107563963 A CN107563963 A CN 107563963A
- Authority
- CN
- China
- Prior art keywords
- resolution
- depth map
- edge
- graph
- input
- 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
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of method based on individual depth map super-resolution rebuilding, it is related to computer vision and image processing field.This method can pass through its high-resolution depth graph of individual depth map reconstruction of acquisition.First, using local self-similarity, the low resolution depth map construction itself training data sample set of input is passed through;Then, using Markov random field model, high-resolution depth edge figure is rebuild by the low resolution depth map of input and the self similarity sample set of construction;Then, in the case where the high-resolution depth edge figure of reconstruction instructs, high-resolution depth graph is recovered by the bilateral filtering of amendment;Finally, processing is iterated using obtained high-resolution depth graph as low resolution input figure, until reaching target resolution.Invention introduces the method that the local self-similarity of image and margin guide depth map recover, and fast and effectively can carry out super-resolution reconstruction to single low-resolution depth map.
Description
Technical field
The present invention relates to computer vision and image processing field, and in particular to one kind is based on individual depth map super-resolution
The method of reconstruction.
Background technology
Depth map is in computer vision fields such as 3DTV, three-dimensional modeling, robot navigation, target tracking, interactive entertainments
In have a wide range of applications.But compared with high-resolution cromogram, the depth map obtained currently with depth camera is differentiated
Rate is very low, and which greatly limits the further use of depth map.For example, Swiss Range SR400 and PMD
The depth map resolution ratio that CamCube is obtained only has 200X200, even Kinect, the depth map resolution ratio got also only has
512X424, far below its corresponding cromogram resolution ratio 1920X1080.Therefore, in order to overcome this problem, depth map is lifted
Resolution ratio becomes a key and urgent research contents.
At present, depth map super-resolution method can be divided into three major types.The first kind is the method based on fusion, this kind of method
It is intended to by merging several low resolution depth maps, to obtain high-resolution depth graph.But this kind of method relies heavily on
In a hypothesis:The still image of multiple scopes can be got by small camera motion, but this should in many reality
It is invalid with middle possibility, and running cost is sufficiently expensive.Second class is based on the method for combining cromogram, by using high score
Structural relation between resolution cromogram and low resolution depth map, to instruct depth map to carry out super-resolution.This kind of method meeting
The problems such as bringing the texture replication of cromogram, also, cromogram and depth map are carried out in a practical situation registering and synchronous
It is a stubborn problem.3rd class is the method based on individual depth map, and this kind of method uses for reference individual natural image super-resolution
Rate method, the method for being based especially on sample learning.But compared to natural image super-resolution, the information that depth map possesses is more
Few, the removal of reservation and noise to edge requires higher, faces more challenges.What but this kind of method need not be extra
Depth image frame or corresponding high-resolution colour picture, therefore be more easily implemented.
The content of the invention
The technical problem to be solved in the present invention is:Overcome the deficiencies in the prior art, there is provided one kind is surpassed based on individual depth map
The method of resolution reconstruction.Guided using local self-similarity and high-resolution edge graph, effectively solve the high score of reconstruction
The problems such as resolution depth map edge ring, obtain preferably rebuilding effect.Realization shows that method proposed by the present invention can be fast
Speed, its high-resolution depth graph is effectively rebuild from individual depth map.
The present invention solve the above problems the technical scheme that uses for:A kind of side based on individual depth map super-resolution rebuilding
Method, realize that step is as follows:
Step (1), training data sample set construction, using the local self-similarity of image, pass through low resolution depth map
Itself, carries out piecemeal processing with reference to Image Edge-Detection operator and shock filtering methods, and by image after processing, is trained
Sample set of data samples.
Step (2), high-resolution edge graph are rebuild, and input sample block is obtained by the low resolution depth map piecemeal of input
It is right, try to achieve the N group training sample blocks pair in (1) with input sample block arest neighbors using range conversion and Euclidean distance.In this base
Markov model is built on plinth, tries to achieve optimal matched sample block, and then obtained final sample block is subjected to fusion treatment
Obtain high-resolution edge graph.
Step (3), depth map super-resolution, high-resolution edge graph is obtained using step (2), with reference to the low resolution of input
Rate depth map, final high-resolution depth graph is obtained using the two-sided filter of an amendment.
Step (4), iterative processing, until reaching target resolution.
Further, the particular content that training data sample set constructs in the step (1) is as follows:
Step (A1), using local self-similarity, the small sample block in image is protected under small change of scale with self structure
Hold similar.Therefore in order to by depth map construction itself self similarity set of data samples, be carried out to the low resolution depth map D of input
The down-sampling of multiple is further set, obtains Dl.Using Image Edge-Detection operator, to DlExtraction edge obtains El, to input
Low resolution depth map extraction edge obtains E, and carries out shock to E and filter to obtain Els。
Step (A2), to edge graph El, EsPiecemeal processing is carried out respectively, respectively obtains sample set of blocks yi lAnd yi sForm instruction
Practice sample block to Y={ yi l, yi s, its effect is similar to external data collection sample block pair.
Further, the particular content that step (2) the middle high-resolution edge graph is rebuild is as follows:
Step (B1), first, it is corresponding defeated with training data sample set in (1) using image local self-similarity, construction
Enter sample data set.The depth map D that low resolution depth map D and further down-sampling to input are obtainedlSet respectively
The up-sampling of multiple obtains DhAnd Dm.Using Image Edge-Detection operator, to DhExtraction edge obtains Eh, to DmExtraction edge obtains
Em, and to EhShock is carried out to filter to obtain Ehs。
Step (B2), to edge graph Em,EhsPiecemeal processing is carried out, respectively obtains sample set of blocks xi lAnd xi sForm input
Sample block is to collecting X={ xi l, xi s, with the training sample block that is obtained in (1) to collecting Y={ yi l, yi sCorresponding;
Step (B3), using input sample block obtained in the previous step to obtained training sample block in collection and (1) to collection, it is first
Range conversion first is carried out to all sample blocks, then tries to achieve the centering of training sample block and each input sample block using Euclidean distance
Adjust the distance nearest N number of training sample block pair.
Step (B4), by each input sample block to Xi, will be obtained in the previous step closest therewith as observer nodes
N training sample block to as hidden node label, building Markov random field model, trying to achieve blocks and optimal matching blocks.
Step (B5), the optimal high-resolution edge samples block that previous step is obtained by Markov model melted
Close, obtain final high-resolution edge graph.
Further, the particular content of depth map super-resolution is as follows in the step (3):
Step (C1), a support window is defined, using the high-resolution edge graph obtained in step (3), to every in figure
Individual pixel, judge whether other pixels belong in edge the same side with current pixel in the support window centered on the pixel.
It is otherwise invalid weight pixel if so, being then used as effective weight pixel.
Step (C2), edge the same side pixel to trying to achieve, it is found corresponding in the low resolution depth map of input
Pixel depth value, and they are weighted ultimate depth value of the summation as high-resolution depth graph.For in support window
Do not belong to the pixel of edge the same side, then pass through phase in depth map after bicubic interpolation with the low resolution depth map of input
Filling should be worth.
Further, the particular content of iterative processing is as follows in the step (4):
Step (D), using the high-resolution depth graph that step (3) obtains as input, return to step (1) is iterated place
Reason, until reaching target depth figure resolution ratio.
It is of the invention compared with present technology the advantages of be:
The present invention draws image local self-similarity and high-resolution edge using single low-resolution depth map as input
Guiding method is combined, and realizes individual depth map super-resolution.Compared with depth map super-resolution method before, method of the invention
Do not need additional depth picture frame or corresponding high-resolution colour picture, it is not required that external data collection, only utilize itself
Information, realize simply, it is good to rebuild effect.
Brief description of the drawings
Fig. 1 is the schematic flow sheet of individual depth map super-resolution of the present invention:
Fig. 2 a~Fig. 2 d are the emulation experiment test image that the present invention uses, respectively in Middlebury data sets
" Cones ", " Teddy ", " Tsukuba " and " Venus " carry out the low resolution depth map after 4 times of down-samplings;
Fig. 3 a~Fig. 3 e are of the invention and existing three kinds of methods to Fig. 2 a Image Reconstruction comparative result figures.Wherein Fig. 3 a are height
Depth of resolution is truly schemed, and Fig. 3 b are the high-resolution depth graph that arest neighbors interpolation method is drawn, Fig. 3 c are the side based on sample
The high-resolution depth graph that method (PB) is drawn, the high-resolution depth graph that Fig. 3 d draw for the method (EG) of margin guide, Fig. 3 e
The high-resolution depth graph drawn for the present invention;
Fig. 4 a~Fig. 4 e are of the invention and existing three kinds of methods to Fig. 2 b image super-resolution comparative result figures, control methods
Same Fig. 3;
Fig. 5 a~Fig. 5 e are of the invention and present three kinds of methods to Fig. 2 c image super-resolution comparative result figures, control methods
Same Fig. 3;
Fig. 6 a~Fig. 6 e are of the invention and present three kinds of methods to Fig. 2 d image super-resolution comparative result figures, control methods
Same Fig. 3;
Fig. 7 a~Fig. 7 d are that the real depth map of the invention collected with existing three kinds of methods to Kinect carries out super-resolution
The comparative result figure of rate.Wherein Fig. 7 a are the high-resolution depth graph that nearest interpolation method is drawn, Fig. 7 b are the side based on sample
The high-resolution depth graph that method (PB) is drawn, the high-resolution depth graph that Fig. 7 c draw for the method (EG) of margin guide, Fig. 7 d
The high-resolution depth graph drawn for the present invention.
Embodiment
The present invention is described in further detail with example below in conjunction with the accompanying drawings:
As shown in figure 1, the invention provides a kind of method based on individual depth map super-resolution rebuilding, including training number
According to sample set constitution step, high-resolution difficulty edge graph reconstruction procedures and depth map super-resolution step.It is of the invention specific real
It is now as follows:
Step 1:Training data sample set constructs, and using local self-similarity, passes through low resolution depth map construction itself
Training data sample set:
In natural image, after image carries out sampling transformation by small decimation factor, fractional sample block can and itself be kept
It is similar.Inspired by this, propose a depth map super-resolution strategy:Pass through the high-resolution depth graph sample up-sampled to small yardstick
The low-frequency component of this block and the low-frequency component of the low resolution depth map sample block of input are matched, and find sample block at low point
Position in resolution figure, then high-resolution depth graph is filled using the radio-frequency component of the low resolution figure opening position.
In order to using local self-similarity construction training sample set, further be set to the low resolution depth map D of input first
The down-sampling processing of multiple, obtains depth map D1.Then using Canny operators to D1Carry out edge extracting and obtain corresponding depth
Edge graph E1.In addition, to the low resolution depth map D of input, edge extracting also is carried out using Canny operators, is obtained corresponding
Depth edge figure E, then shock filtering process, the depth edge figure E sharpened are carried out to Els。
Then, the depth edge figure E to obtainingl,EsPiecemeal processing is carried out respectively, respectively obtains sample set of blocks { yi lAnd
{yi s, composing training sample block is to Y={ yi l, yi s}.Herein, being divided into for sample block is extracted pixel-by-pixel, i.e., from the image upper left corner
To the image lower right corner, the extraction of sample block is carried out centered on each pixel, wherein, ysBlock size is ylDecimation factor multiple
Add 1, such as:ylBlock size is 3*3, and the decimation factor set is 2, then corresponding ysBlock size is 7*7.
Step 2:High-resolution depth edge figure is rebuild, using being obtained in the low resolution depth map and step 1 of input
Training sample block to collection, construct Markov random field model:
First, input sample block collection corresponding with training sample block collection in (1) is constructed.To the low resolution depth map of input
The depth map D that D and further down-sampling obtainlThe up-sampling for carrying out setting multiple respectively obtains DhAnd Dm, herein on adopt
Quadrat method is bicubic interpolation method.Canny operators are recycled, to DhExtraction edge obtains corresponding depth edge figure Eh, use
Same method is to DmProcessing obtains depth edge figure Em, and to EhCarry out shock and filter the depth edge figure E sharpenedhs。
Then, to depth edge figure Em, EhsPiecemeal processing is carried out respectively, respectively obtains sample set of blocks xi lAnd xi sForm
Input sample block is to collecting X={ xi l, xi s, with the training sample block that is obtained in (1) to collecting Y={ yi l, yi sCorresponding.Pay attention to, herein
Piecemeal size and step 1 sample block are in the same size.But sample is extracted pixel-by-pixel when being constructed different from training sample block in step 1
Block, the nonoverlapping image pattern block for being here divided into whole image.
Then, to each input sample block pair, as the observer nodes in Markov random field, and with European
Distance tries to achieve training sample block to concentrating the N number of training sample block pair closest with it, then by this N number of training sample block pair
As the hidden node label in Markov random field, Markov random field model is constructed.And utilize the random mould of markov
Type tries to achieve blocks and optimal matching blocks.Wherein, markov energy function is as follows:
Wherein, E1And E2For data item, E3For smooth item, β and γ are respectively weight coefficient.First data item E1Weigh
Training sample block yi lWith input sample block xi lBetween similitude;Second data item E2Weigh training sample block yi sWith input
Sample block xi sBetween similitude..Wherein:
Here, d is the range conversion of edge samples block, can be to two by carrying out Euclidean distance calculating after range conversion again
Value pattern carries out more preferable similarity measurement.
Smooth item E3Strengthen the uniformity of neighboring edge sample block overlapping region, wherein OijIt is to neighboring edge sample block
yi s Withyj sOverlapping region carries out the operation of extracted region.
By horse can husband's model optimal solution is tried to achieve to energy function, the optimal high-resolution edge samples that will finally obtain
Block is merged, and obtains final high-resolution depth edge figure.
Step 3:Depth map super-resolution, using high-resolution depth edge figure, with reference to the low resolution depth of input
Figure, final high-resolution depth graph is obtained using the two-sided filter of amendment:
Wherein, DhFor target high-resolution depth map, DlFor the low resolution depth map of input, EhFor high-resolution edge
Figure, N (p) are the support window centered on pixel p of definition, and p ↓ and q ↓ is pixel in the low resolution depth map of input, kp
It is normalization factor, fd(*) is gaussian kernel function, fr(*) is binary indicator function, is defined as follows:
By the guiding of high-resolution depth edge figure, only the pixel in edge the same side, which can be weighted, remains into finally
High-resolution depth graph in.Belong to the pixel of edge the same side pixel for being not present in support window, then with input
Low resolution depth map is filled by corresponding depth value in the depth map that is obtained after bicubic interpolation.
As shown in table 1, compared by evaluation index and the existing method that represents of root-mean-square error RMSE:Arest neighbors
Method (PB, 2012) based on sample that the method (NN), O.M.Aodha etc. 12 years of interpolation proposes and Jun Xie etc. were in 14 years
It is proposed the method (EG, 2014) of margin guide.Method proposed by the present invention based on individual depth map super-resolution rebuilding can
Obtain the result compared with high realism and accuracy.Visual effect results contrast such as Fig. 3,4,5,6,7.
The quantitative comparison of different super-resolution methods on table 1Middlebury data sets
Cones | Venus | Teddy | Tsukuba | |
NN | 1.498 | 0.367 | 1.348 | 0.832 |
PB | 1.481 | 0.337 | 1.280 | 0.833 |
EG | 1.157 | 0.314 | 1.024 | 0.765 |
The present invention | 1.052 | 0.247 | 0.888 | 0.723 |
The content not being described in detail in description of the invention belongs to prior art known to professional and technical personnel in the field.
Described above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these improvements and modifications also should
It is considered as protection scope of the present invention.
Claims (5)
- A kind of 1. method based on individual depth map super-resolution rebuilding, it is characterised in that this method step is as follows:Step (1), training data sample set construction, using local self-similarity, pass through the low resolution depth map itself of input Construct training data sample set;Step (2), high-resolution edge graph rebuild, using input low resolution edge graph and construction self similarity sample set, High-resolution edge graph is rebuild by Markov random field model;Step (3), depth map super-resolution, it is extensive by the bilateral filtering of amendment in the case where the high-resolution edge graph of reconstruction instructs Multiple high-resolution depth graph;Step (4), iterative processing, using obtained high-resolution depth graph as input iterative processing, differentiated until reaching target Rate.
- A kind of 2. method based on individual depth map super-resolution rebuilding according to claim 1, it is characterised in that:It is described The particular content of training data sample set construction in step (1) is as follows:Step (A1), the low resolution depth map D to input carry out the down-sampling of further setting multiple, obtain Dl, utilize image Edge detection operator, to DlExtraction edge obtains edge graph El, in addition, being obtained to the low resolution depth map extraction edge of input E, and shock is carried out to E and filters to obtain Els;Step (A2), to edge graph El, ElsPiecemeal processing is carried out respectively, respectively obtains sample set of blocks yi lAnd yi sComposing training sample This block is to collecting Y={ yi l, yi s}。
- A kind of 3. method based on individual depth map super-resolution rebuilding according to claim 1, it is characterised in that:It is described The particular content that high-resolution edge graph in step (2) is rebuild is as follows:Step (B1), the depth map D obtained to the low resolution depth map D of input and further down-samplinglCarry out setting again respectively Several up-samplings obtain DhAnd Dm, using Image Edge-Detection operator, to DhExtraction edge obtains Eh, to DmExtraction edge obtains Em, and to EhShock is carried out to filter to obtain Ehs;Step (B2), to edge graph Em,EhsPiecemeal processing is carried out respectively, respectively obtains sample set of blocks xi lAnd xi sForm input sample This block is to collecting X={ xi l, xi s};Step (B3), to each input sample block pair, it is European to carrying out with all training sample blocks for being obtained in (1) respectively Distance calculates, and is adjusted the distance nearest N number of training sample block pair with each input sample block with trying to achieve;Step (B4), by each input sample block to XiAs observer nodes, that is tried to achieve in previous step is closest therewith N number of Training sample block builds Markov random field model as hidden node label;Step (B5), the high-resolution edge samples block obtained by Markov random field model merged, obtained final High-resolution edge graph.
- A kind of 4. method based on individual depth map super-resolution rebuilding according to claim 1, it is characterised in that:It is described The particular content of depth map super-resolution in step (3) is as follows:Step (C1), a support window is defined, using the high-resolution edge graph tried to achieve in step (3), to each picture in figure Element, judge that whether other pixels and current pixel are in the same side at edge in the support window centered on the pixel;Step (C2), the pixel positioned at edge the same side to trying to achieve, find its pixel depth in low resolution depth map Value, and they are weighted ultimate depth value of the summation as high-resolution depth graph.
- A kind of 5. method based on individual depth map super-resolution rebuilding according to claim 1, it is characterised in that:It is described The particular content of iterative processing in step (4) is as follows:Step (D), using the high-resolution depth graph that step (3) obtains as input, return to step (1) is iterated processing, directly Reach target depth figure resolution ratio.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710686263.6A CN107563963B (en) | 2017-08-11 | 2017-08-11 | Super-resolution reconstruction method based on single depth map |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710686263.6A CN107563963B (en) | 2017-08-11 | 2017-08-11 | Super-resolution reconstruction method based on single depth map |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107563963A true CN107563963A (en) | 2018-01-09 |
CN107563963B CN107563963B (en) | 2020-01-03 |
Family
ID=60975395
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710686263.6A Active CN107563963B (en) | 2017-08-11 | 2017-08-11 | Super-resolution reconstruction method based on single depth map |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107563963B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107680042A (en) * | 2017-09-27 | 2018-02-09 | 杭州群核信息技术有限公司 | Rendering intent, device, engine and storage medium |
CN109731238A (en) * | 2019-01-10 | 2019-05-10 | 李荻 | Pattern switching platform based on site environment |
CN111489383A (en) * | 2020-04-10 | 2020-08-04 | 山东师范大学 | Depth image up-sampling method and system based on depth edge point and color image |
CN112308781A (en) * | 2020-11-23 | 2021-02-02 | 中国科学院深圳先进技术研究院 | Single image three-dimensional super-resolution reconstruction method based on deep learning |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102142137A (en) * | 2011-03-10 | 2011-08-03 | 西安电子科技大学 | High-resolution dictionary based sparse representation image super-resolution reconstruction method |
CN103020897A (en) * | 2012-09-28 | 2013-04-03 | 香港应用科技研究院有限公司 | Device for reconstructing based on super-resolution of multi-block single-frame image, system and method thereof |
CN104766273A (en) * | 2015-04-20 | 2015-07-08 | 重庆大学 | Infrared image super-resolution reestablishing method based on compressed sensing theory |
-
2017
- 2017-08-11 CN CN201710686263.6A patent/CN107563963B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102142137A (en) * | 2011-03-10 | 2011-08-03 | 西安电子科技大学 | High-resolution dictionary based sparse representation image super-resolution reconstruction method |
CN103020897A (en) * | 2012-09-28 | 2013-04-03 | 香港应用科技研究院有限公司 | Device for reconstructing based on super-resolution of multi-block single-frame image, system and method thereof |
CN104766273A (en) * | 2015-04-20 | 2015-07-08 | 重庆大学 | Infrared image super-resolution reestablishing method based on compressed sensing theory |
Non-Patent Citations (2)
Title |
---|
MICHAEL HORNACEK 等: "Depth Super Resolution by Rigid Body Self-Similarity in 3D", 《THE IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
XIAOCHUAN WANG 等: "Single Depth Map Super-resolution with Local Self-similarity", 《ICVIP 2018 PROCEEDINGS OF THE 2018 THE 2ND INTERNATIONAL CONFERENCE ON VIDEO AND IMAGE PROCESSING》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107680042A (en) * | 2017-09-27 | 2018-02-09 | 杭州群核信息技术有限公司 | Rendering intent, device, engine and storage medium |
CN107680042B (en) * | 2017-09-27 | 2020-03-31 | 杭州群核信息技术有限公司 | Rendering method, device, engine and storage medium combining texture and convolution network |
CN109731238A (en) * | 2019-01-10 | 2019-05-10 | 李荻 | Pattern switching platform based on site environment |
CN111489383A (en) * | 2020-04-10 | 2020-08-04 | 山东师范大学 | Depth image up-sampling method and system based on depth edge point and color image |
CN111489383B (en) * | 2020-04-10 | 2022-06-10 | 山东师范大学 | Depth image up-sampling method and system based on depth marginal point and color image |
CN112308781A (en) * | 2020-11-23 | 2021-02-02 | 中国科学院深圳先进技术研究院 | Single image three-dimensional super-resolution reconstruction method based on deep learning |
Also Published As
Publication number | Publication date |
---|---|
CN107563963B (en) | 2020-01-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107194872B (en) | Remote sensed image super-resolution reconstruction method based on perception of content deep learning network | |
CN104134234B (en) | A kind of full automatic three-dimensional scene construction method based on single image | |
CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
CN107563963A (en) | A kind of method based on individual depth map super-resolution rebuilding | |
CN107563438A (en) | The multi-modal Remote Sensing Images Matching Method and system of a kind of fast robust | |
CN103208102B (en) | A kind of remote sensing image fusion method based on rarefaction representation | |
Hua et al. | A normalized convolutional neural network for guided sparse depth upsampling. | |
Qin et al. | Bylabel: A boundary based semi-automatic image annotation tool | |
CN102930518B (en) | Improved sparse representation based image super-resolution method | |
CN103761739B (en) | A kind of Image registration method optimized based on half global energy | |
CN105488776B (en) | Super-resolution image reconstruction method and device | |
CN108090403A (en) | Face dynamic identification method and system based on 3D convolutional neural network | |
CN106339998A (en) | Multi-focus image fusion method based on contrast pyramid transformation | |
CN110827312B (en) | Learning method based on cooperative visual attention neural network | |
CN104881662A (en) | Single-image pedestrian detection method | |
CN106295637A (en) | A kind of vehicle identification method based on degree of depth study with intensified learning | |
CN106920214A (en) | Spatial target images super resolution ratio reconstruction method | |
CN112489050A (en) | Semi-supervised instance segmentation algorithm based on feature migration | |
CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image | |
CN107274378A (en) | A kind of image blurring type identification and parameter tuning method for merging memory CNN | |
CN107680140A (en) | A kind of depth image high-resolution reconstruction method based on Kinect cameras | |
CN110189390A (en) | A kind of monocular vision SLAM method and system | |
CN101635048A (en) | Super-resolution processing method of face image integrating global feature with local information | |
CN106169174A (en) | A kind of image magnification method | |
CN107527054A (en) | Prospect extraction method based on various visual angles fusion |
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 |