CN105574876B - A kind of natural image dividing method based on label semantic reasoning - Google Patents
A kind of natural image dividing method based on label semantic reasoning Download PDFInfo
- Publication number
- CN105574876B CN105574876B CN201510957665.6A CN201510957665A CN105574876B CN 105574876 B CN105574876 B CN 105574876B CN 201510957665 A CN201510957665 A CN 201510957665A CN 105574876 B CN105574876 B CN 105574876B
- Authority
- CN
- China
- Prior art keywords
- image
- pixel
- label
- subgraph
- region
- 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 Analysis (AREA)
Abstract
A kind of natural image dividing method based on label semantic reasoning of the disclosure of the invention, belongs to digital image processing field, is related to image segmentation task.A kind of natural image automatic division method of proposition of the present invention, this method marks the cut zone of image using numeral, image-region number after automatic estimation segmentation, so that one independent object of each piece of region representation, and continuity is not lost between region, so as to more meet the understanding demand of human perception.On the premise of ensureing the strong association even between the strong similitude and cut zone in cut zone, more significant image segmentation result is obtained.
Description
Technical field
The invention belongs to digital image processing field, is related to image segmentation task.
Background technology
Image segmentation refers to divide the image into some specific, the technology and process of the target area with certain property.
It is computer vision, in image processing field the problem of the bottom.Image segmentation can be used for target detection, identification and image
Understand in contour level Computer Vision Task.According to the difference for the task that is applicable, there has been proposed a variety of research methods：
(1) image partition method based on cluster
This method is intended to by maximizing the non-similarity between class and minimizing the similitude in class, so as to which image be divided
For different clusters.The algorithm of comparative maturity is including K-means, Mean-shift, and Quick-shift etc..
(2) image partition method based on graph theory
The thought of this method is to regard each pixel in image as a node, the connection weight between two nodes
The similarity being proportional between pixel pair, such sub-picture are converted to a full connection figure, so as to which segmentation problem be changed
For Graph partition problem.The method of comparative maturity includes N-cut, Egbis etc..
(3) image partition method based on layering
This method is method popular in recent years, and it focuses on designing different hierarchical modes to integrate various differences
The information such as the color of image of yardstick, texture, one are formed by the thick structure to essence, so as to reach the purpose of image segmentation.
And in natural image segmentation, its difficult point is：Target is relatively easy in image, causes same target to be divided
For some different regions；Or target is excessively complicated in image, cause the meaning expressed by cut zone not strong.In computer
In vision in most of tasks, if single Object Segmentation does not combine with mark, it is difficult to form the feature of semanteme, this is in high meter
Lack Practical significance in calculation machine visual task.
The content of the invention
In order to overcome traditional images dividing method due between the image-region after segmentation relevance it is not strong caused by people
Class perceive deviation deficiency, the present invention have studied how ensure cut zone in strong similitude and cut zone between strong pass
On the premise of connection connects, more significant image segmentation result is obtained.
It is an object of the invention to propose a kind of natural image automatic division method, this method utilizes numeral mark image
Cut zone, the image-region number after automatic estimation segmentation so that one independent object of each piece of region representation, and area
Continuity is not lost between domain, so as to more meet the understanding demand of human perception.
The present invention devises entitled LSI (Latent Semantic Inference of Serial Label)
Image segmentation framework.The framework includes two stages：Label generation phase and label reasoning stage.
Thus a kind of natural image dividing method based on label semantic reasoning of the present invention, this method include:
Step 1:An image to be split is obtained, four subgraphs will be divided into segmentation image averaging；
Step 2：Each subgraph is handled as follows；
Step 2.1：Calculate the Euclidean distance of each pixel between any two in subgraph；
Step 2.2：Calculate the average Euclidean distance between each pixel and remaining all pixels point in subgraph；
Step 2.3：The local density of each pixel is calculated, the local density represents certain pixel to the Europe of rest of pixels point
Family name's distance is more than the number of average Euclidean distance；
Step 2.4：For a certain pixel, the rest of pixels point that local density is more than the pixel is found, finds these
The minimum value of the pixel Euclidean distance is arrived in pixel, the minimum value is the similarity distance for pixel；
Step 2.5：The similarity distance of all pixels point and neighbours in subgraph are obtained using the method for step 2.4；
Step 2.6：2-6 pixel of local density and similarity distance maximum in subgraph is chosen, these pixels are
For the cluster centre of the subgraph, the subgraph is divided into corresponding region according to these cluster centres, and to each region
Add label；
Step 3：Each subgraph for having divided region is spliced, the region that abutment joint both sides will splice is known
Not, whether judge that the two belong to the region of subgraph in original image is the same area；
Step 3.1：Count the number N of pixel at two area joints that will splicepair；
Step 3.2：Calculate the fusion distance S in the region that will splicepair, i.e. seam crossing pixel pair between two regions
Euclidean distance sum；
Step 3.3：Calculate the similarity distance sum for the 2 region all pixels points that will splice；
Step 3.4：If Npair＞ 5 and SpairSimilarity distance sum is obtained then by two regions to be spliced less than step 3.3
It is the same area in splicing backsight, and adds label again；
Step 3.4：Each subgraph is stitched together using the method for step 3.1-3.3；
Step 4：Step 1 is obtained into image to be split, carries out size reduction twice respectively, it is different small to obtain two sizes
Image；
Step 5：Using step 2 neutron image processing method, region division is carried out to each small image, and each region is added
Label；
Step 6：All label areas of two small images of the tape label that image after splicing and step 5 are obtained are carried out pair
Than if having the label area of two images consistent in three images, the label area of a remaining image is replaced with
Label as two images before, in this process, if the area label of image changes after splicing, again
Image mosaic is carried out until obtaining the final segmentation result of input picture using step 3.
The present invention has advantages below：
(1) divide the image into problem and be converted to label reasoning problems.
(2) this method can automatically estimating picture segmentation after number of regions, to meet the understanding demand of human perception；
(3) this method devises an image segmentation framework (LSI)；In LSI frameworks, the design of this method obtains in detail
Thin image local information；The global information for incorporating image reasonable in design of tag fusion strategy；Layered label alignment machine
The effective dimensional information that make use of image of design of system.
Brief description of the drawings
Fig. 1 is system framework figure of the present invention.
Fig. 2 is the natural image dividing method result schematic diagram of the present invention based on label semantic reasoning.
Fig. 3 is the natural image segmentation result schematic diagram obtained by the present invention.
Embodiment
In order to solve the above problems, the natural image dividing method proposed by the present invention based on label semantic reasoning is specifically real
It is as follows to apply step：
Step 1：Input picture is divided into four subgraphs along two medium lines, respectively to each subgraph
Perform following operation；
(1) positional information of the colouring information to CIELAB spaces and each pixel, Euclidean distance formula meter is passed through
Calculate the distance between all pixels pair in each subgraph；
(formula 1)
Wherein, (l, a, b, x, y) is the vector representation of each pixel of CIELAB spaces,It is two pixel pi,
pjPixel pair between distance.
(2) using the distance between the obtained pixel pair in previous step, the local density of each pixel is calculated, simultaneously
The distance of each pixel and other pixels more than the pixel local density is calculated, takes the minimum value in all distances
The as similarity distance of the pixel, corresponding pixel are the neighbours of the pixel；
(formula 2)
Wherein, χ (x)=1, ifx<0, or χ (x)=0, otherwise.T isAverage.It is pixel piOffice
Portion's density.
(formula 3)
Wherein,It is pixel piSimilarity distance.Formula 3 meets condition
(3) using with top-down mode, choosing in subgraph local density and similarity distance relatively most from right to left
2-6 big pixel, these pixels are the cluster centre of subgraph, each cluster corresponding to attached pixel from its
Found in corresponding neighbor pixel point.Cluster centre in subgraph shown in the form of digital label, as next stage
Input label；
Step 2：Input label is modeled, the energy function of Cross-R＆S models is as follows：
Wherein,For unitary item, φijFor binary item, M is the number of input label, and N is the number of neighborhood label.
Step 3：Utilize tag fusion policy update, it is tactful as follows；
(1) cluster centre (or referred to as subregion) obtained for label generation phase, using following strategy
(2) along medium line, the number N of neighbor pixel pair between different zones is calculatedpairInterregional fusion two-by-two
Distance Spair；If Npair＞ 5 andIt is one by two Cluster mergings, whereinRepresent son
Each zonule in figure, i represent the sequence number of subgraph, and j represents the number of cluster centre, i.e. two sub-regions merge into one newly
Region, until along all cluster centre numbers of medium line being kept to 1 by total k.
(3) according to the strategy, updated using formula 5；
Step 4：Utilize label registration new mechanism φij, mechanism is as follows
(1) input picture being carried out into average twice will sample, and two scalogram pictures be obtained, with CLDC algorithms to this two width chi
Degree image is handled, and thus, we, which amount to, obtains three width label images, is respectively designated as S1-1, S2-1, and S3-1, wherein,
S3-1 metric space layout information is more than S2-1 and S1-1.
(2) for S3-1, S1-1 and S2-1 are updated according to its subtab respectively；If S1-1 subtabs change,
SpairValue can produce change, now need again use step 3 in tag fusion strategy to S1-1 carry out tag fusion,
Obtain new S1-1', now, for S2-1, if S1-1' and S3-1 label in same position and S2-1 label not
Together, then be replaced the label of relevant position in S2-1, otherwise, the label in S2-1 is constant.Thus, new label is obtained
Image S2-1'.
(3) according to the strategy, φ is updated using formula 6ij；
(formula 6)
Wherein, NumS3-1-NumS2-1It is the pixel of label image S3-1 and S2-1 different labels in alignment mechanism each time
Count out.
Step 5：Iteration is updated to E convergences, and we can obtain final segmentation figure picture.
Claims (1)
1. a kind of natural image dividing method based on label semantic reasoning, this method include:
Step 1:An image to be split is obtained, image averaging to be split is divided into four subgraphs；
Step 2：Each subgraph is handled as follows：
Step 2.1：Calculate the Euclidean distance of each pixel between any two in subgraph；
Step 2.2：Calculate the average Euclidean distance between each pixel and remaining all pixels point in subgraph；
Step 2.3：Calculate the local density of each pixel, the local density represent certain pixel to rest of pixels point Euclidean away from
From the number more than average Euclidean distance；
Step 2.4：For a certain pixel, the rest of pixels point that local density is more than the pixel local density is found, is found
The minimum value of the pixel Euclidean distance is arrived in these pixels, the minimum value is the similarity distance for pixel；
Step 2.5：The similarity distance of all pixels point and neighbours in subgraph are obtained using the method for step 2.4；
Step 2.6：2-6 pixel of local density and similarity distance maximum in subgraph is chosen, these pixels are should
The cluster centre of subgraph, the subgraph is divided into by corresponding region according to cluster centre, and label is added to each region；
Step 3：Each subgraph for having divided region is spliced, the region that abutment joint both sides will splice is identified, and sentences
Whether the two disconnected regions for belonging to subgraph are the same area in original image；
Step 3.1：Count the number N of pixel at two area joints that will splicepair；
Step 3.2：Calculate the fusion distance S in two regions that will splicepair, i.e. the Europe of seam crossing pixel pair between two regions
Formula is apart from sum；
Step 3.3：Calculate the similarity distance sum for the 2 region all pixels points that will splice；
Step 3.4：If Npair＞ 5 and SpairSimilarity distance sum is obtained less than step 3.3 then to spell two regions to be spliced
It is the same area to connect backsight, and adds label again；
Step 3.5：Each subgraph is stitched together using the method for step 3.1-step 3.4；
Step 4：Step 1 is obtained into image to be split, carries out size reduction twice respectively, obtains the different small figure of two sizes
Picture；
Step 5：Using step 2 neutron image processing method, region division is carried out to each small image, and each region is added and marked
Label；
Step 6：All label areas of two small images of the tape label that image after splicing and step 5 are obtained are contrasted,
If having the label area of two images consistent in three images, the label area of a remaining image is replaced with and marked
The same label of two consistent images of region is signed, in this process, if the area label of image changes after splicing,
Image mosaic is then carried out until obtaining the final segmentation result of input picture using step 3 again.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510957665.6A CN105574876B (en) | 2015-12-18 | 2015-12-18 | A kind of natural image dividing method based on label semantic reasoning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510957665.6A CN105574876B (en) | 2015-12-18 | 2015-12-18 | A kind of natural image dividing method based on label semantic reasoning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105574876A CN105574876A (en) | 2016-05-11 |
| CN105574876B true CN105574876B (en) | 2018-02-09 |
Family
ID=55884967
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510957665.6A Expired - Fee Related CN105574876B (en) | 2015-12-18 | 2015-12-18 | A kind of natural image dividing method based on label semantic reasoning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105574876B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107330520A (en) * | 2017-06-09 | 2017-11-07 | 上海电力学院 | The object Affording acquisition inference method that a kind of knowledge based storehouse is represented |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103268635A (en) * | 2013-05-15 | 2013-08-28 | 北京交通大学 | Segmentation and semantic annotation method of geometry grid scene model |
| CN103945227A (en) * | 2014-04-16 | 2014-07-23 | 上海交通大学 | Video semantic block partition method based on light stream clustering |
| US20150206315A1 (en) * | 2014-01-21 | 2015-07-23 | Adobe Systems Incorporated | Labeling Objects in Image Scenes |
-
2015
- 2015-12-18 CN CN201510957665.6A patent/CN105574876B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103268635A (en) * | 2013-05-15 | 2013-08-28 | 北京交通大学 | Segmentation and semantic annotation method of geometry grid scene model |
| US20150206315A1 (en) * | 2014-01-21 | 2015-07-23 | Adobe Systems Incorporated | Labeling Objects in Image Scenes |
| CN103945227A (en) * | 2014-04-16 | 2014-07-23 | 上海交通大学 | Video semantic block partition method based on light stream clustering |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105574876A (en) | 2016-05-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | PMSC: PatchMatch-based superpixel cut for accurate stereo matching | |
| Dong et al. | An efficient global energy optimization approach for robust 3D plane segmentation of point clouds | |
| Mansouri et al. | Multiregion competition: A level set extension of region competition to multiple region image partitioning | |
| CN108961327A (en) | A kind of monocular depth estimation method and its device, equipment and storage medium | |
| Idrees et al. | Tracking in dense crowds using prominence and neighborhood motion concurrence | |
| Ni et al. | Learning to photograph: A compositional perspective | |
| Zhao et al. | Learning deep network for detecting 3d object keypoints and 6d poses | |
| CN103400386A (en) | Interactive image processing method used for video | |
| CN102360504A (en) | Self-adaptation virtual and actual three-dimensional registration method based on multiple natural characteristics | |
| CN107909079A (en) | One kind collaboration conspicuousness detection method | |
| Stekovic et al. | General 3d room layout from a single view by render-and-compare | |
| Salas et al. | Layout aware visual tracking and mapping | |
| Yaman et al. | An iterative adaptive multi-modal stereo-vision method using mutual information | |
| CN104331890A (en) | Method and system for estimating global disparity | |
| CN116385660A (en) | Indoor single view scene semantic reconstruction method and system | |
| Prasad et al. | Epipolar geometry based learning of multi-view depth and ego-motion from monocular sequences | |
| Peng et al. | Stereo matching with space-constrained cost aggregation and segmentation-based disparity refinement | |
| CN111402429B (en) | Scale reduction and three-dimensional reconstruction method, system, storage medium and equipment | |
| Yang et al. | Cost-efficient image semantic segmentation for indoor scene understanding using weakly supervised learning and BIM | |
| CN105574876B (en) | A kind of natural image dividing method based on label semantic reasoning | |
| Wang et al. | Tc-sfm: Robust track-community-based structure-from-motion | |
| Dong et al. | Weighted triplet loss based on deep neural networks for loop closure detection in VSLAM | |
| CN103942779A (en) | Image segmentation method based on combination of graph theory and semi-supervised learning | |
| El Ansari et al. | A new regions matching for color stereo images | |
| Zaheer et al. | Shape from angle regularity |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| 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: 20180209 Termination date: 20201218 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |