CN107103579A - A kind of RANSAC improved methods towards image mosaic - Google Patents
A kind of RANSAC improved methods towards image mosaic Download PDFInfo
- Publication number
- CN107103579A CN107103579A CN201710236500.9A CN201710236500A CN107103579A CN 107103579 A CN107103579 A CN 107103579A CN 201710236500 A CN201710236500 A CN 201710236500A CN 107103579 A CN107103579 A CN 107103579A
- Authority
- CN
- China
- Prior art keywords
- point
- image
- ransac
- probability
- pixel
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 238000005070 sampling Methods 0.000 claims abstract description 11
- 239000013598 vector Substances 0.000 claims description 18
- 238000000605 extraction Methods 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 230000001186 cumulative effect Effects 0.000 claims description 7
- 238000005315 distribution function Methods 0.000 claims description 6
- 238000007500 overflow downdraw method Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 238000003708 edge detection Methods 0.000 claims description 3
- 238000005286 illumination Methods 0.000 claims description 3
- 230000004044 response Effects 0.000 claims description 3
- 238000013316 zoning Methods 0.000 claims description 3
- 238000007689 inspection Methods 0.000 claims description 2
- 238000005303 weighing Methods 0.000 claims 1
- 238000009499 grossing Methods 0.000 abstract description 2
- 230000009466 transformation Effects 0.000 description 7
- 238000012952 Resampling Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/14—Transformations for image registration, e.g. adjusting or mapping for alignment of images
- G06T3/153—Transformations for image registration, e.g. adjusting or mapping for alignment of images using elastic snapping
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of RANSAC improved methods towards image mosaic, it is related to computer vision field.This method is by extracting image characteristic point, by the thought of SURF (Speeded Up Robust Features) algorithm, based on Sobel operators, when rejecting is mismatched, using improved probability random sampling uniformity (RANSAC) algorithm, by constantly adjusting sampled probability, so as to obtain correct model with less time, efficiency of algorithm is effectively improved.In image co-registration, luminance proportion pretreatment is first carried out, then is merged with weighting smoothing algorithm, so as to eliminate the splicing line and intermediate zone of result figure, image mosaic quality is improved.
Description
Technical field
The present invention relates to computer vision field, a kind of RANSAC improved methods towards image mosaic are refered in particular to.
Background technology
Image mosaic is the study hotspot of computer vision and image processing field, it be several are existed each other it is overlapping
Partial image sequence carries out spatial match alignment, a width is formed after being merged through resampling comprising each image sequence information, wide
The new images of visual angle scene, high-resolution.
Image mosaic process has 3 key steps：Image preprocessing, images match and image co-registration, wherein images match
It is a most important step.SURF (Speeded-Up Robust Features) algorithm is to use widest characteristic matching
One of algorithm, but after SURF extracts characteristic point, it is necessary to Mismatching point is rejected, probability random sampling uniformity is commonly used
(RANSAC) Mismatching point rejecting is carried out.RANSAC is kind of the mathematical modeling that parameter is estimated using iterative method, with traditional
RANSAC calculating parameters carry out essence with time, and iterations is more, and computing is more time-consuming, have impact on the efficiency of image mosaic.Tradition
The RANSAC algorithms characteristic point selected in initialization model be completely random, not certain constraint rule is led
Cause the real-time reduction of algorithm.
The content of the invention
The technical problem to be solved in the present invention is a kind of quality that can effectively improve image mosaic of proposition and splicing speed
The RANSAC improved methods towards image mosaic of rate.
In order to solve the above technical problems, technical scheme specifically includes following steps：
The first step：Feature point extraction and the generation of description；
Second step：Characteristic matching；
3rd step：Image registration is carried out using improved RANSAC methods；
4th step：Brightness pretreatment is carried out to input picture using histogram equalization method, input image lightness is reached
To equilibrium, then reuse weighted average fusion method and carry out image co-registration.
As the preferred of technical solution of the present invention, feature point extraction specifically includes following steps described in the first step：
(1.1) Sobel operators carry out convolutional calculation using 3 × 3 module, and its Jacobian matrix is
(1.2) useWithTwo convolution masks, utilize this two volumes
Product module plate calculates transverse edge detection G respectivelyxG is detected with longitudinal edgey：
(1.3) by Gx, GyIt can calculate
Then the threshold value set in advance and G (x, y) are compared, if threshold value is less than G (x, y), the then pixel (i, j)
For marginal point, the marginal point is extracted as characteristic point.
As the preferred of technical solution of the present invention, the generation that son is described described in the first step comprises the following steps：
(2.1) centered on the characteristic point of extraction, pair radius is any point in 6 σ circle domain, seeks any in the circle domain
The haar small echo response coefficients of point in the x and y direction；
(2.2) distance of any point in the round domain to feature dot center is a radius, then with a radius
For starting point, the sector region for taking 50 ° is zoning, and it is square counterclockwise in 6 σ circle domain in the radius then to make the sector region
To rotation, to x and y directions haar wavelet coefficients in this sector region and calculated respectively every 30 °；
(2.3) the haar wavelet coefficients and form a vector, a total of 12 vectors determine mould in this 12 vectors
A maximum vector, then using a maximum vectorial direction of the mould as characteristic point principal direction；
(2.4) and then a characteristic point detected by Step1 is chosen, and centered on this feature point, coordinate system is revolved
The principal direction is gone to, and makees the square area that a length of side is 20 σ, the square area is divided into 4 × 4 zonule,
The haar wavelet coefficients in the x and y directions of each zonule pixel are calculated respectively, and the length of side of each zonule when calculating is
The σ of 5 σ × 5, direction is consistent with principal direction；
(2.5) the haar wavelet coefficients in the x and y directions of each pixel in the square area for trying to achieve step (4)
Sum operation is carried out, d is respectively obtainedxAnd dy, then by the absolute of the haar wavelet coefficients of the every bit pixel of the square area
Value summation, respectively obtains ∑ dx, ∑ dy；
(2.6) a vector is obtained in each zonule
V=[∑ dx,∑dy,∑|dx|,∑|dy|], the square area includes 16 zonules, 16 zonules
Window vector constitutes description, and its dimension size is 4 × 4 × 4=64.
As the preferred of technical solution of the present invention, in the second step, in the matching process, phase is carried out using Euclidean distance
Like degree measurement, i.e., with minimum distance with the ratio between time closely, if the ratio is less than some given threshold, then it is assumed that be optimal
Match somebody with somebody, otherwise abandon；
As the preferred of technical solution of the present invention, in the 3rd step, by improving the process of stochastical sampling, using each
Model parameter is estimated in the minimum subsample of extraction, and the model parameter further according to estimation adjusts subsample and is chosen in sampling below
The probability taken, and be constantly iterated to extract to examine and obtain the probability of correct temporary pattern to improve.
As the preferred of technical solution of the present invention, in the 3rd step, data point sum when defining image registration is n,
Data point sample weight isInitial smallest subset is ψ0, during estimation model parameter, initial data is missed
Difference is ei(i=1,2 ..., n), sampled probability is：
P probability threshold values are, it is interim in point set S, (Si,Sj) be arbitrfary point pair in S, its Euclidean distance d=| | Si-Sj| |, D
Euclidean distance threshold value；
Improve RANSAC algorithm flows specific as follows：
Step3.1：Initialization data point sampling weight isSelect initial smallest subset
ψ0；
Step3.2：Computation model parameter and error ei(i=1,2 ..., n)；
Step3.3：Update sample weights ωi, obtain sampled probability p (xi)；
Step3.4：According to sampled probability p (xi), obtain new data subset ψ0；
Step3.5：If p (xi) ＞ P, then Step3.1 is returned, otherwise, that is, temporary pattern is found and obtains point set in interim
S；
Step3.6：If D ＜ d, the final interior point set and final model parameter for meeting model are obtained, terminates algorithm；If d
＞ D, then return to Step3.5 and continue iteration, chooses in interim in point set S arbitrfary point to (Si,Sj), until D ＜ d, acquisition meets mould
The final interior point set of type and final model parameter, terminate algorithm.
As the preferred of technical solution of the present invention, the 4th step specifically includes following steps：
Step4.1：Overlay region image histogram is normalized, the probability density function of overlay region image pixel value is obtained
Formula is equalized, the cumulative distribution function of overlay region pixel value is as follows：
In formula, ω is aleatory variable, and p (ω) is the probability that each pixel value occurs；
Step4.2：According to formula (7), input image pixels value mapping relations are set up, each pixel to image is reflected
Penetrate adjustment；Definition input picture is I1、I2, take I1、I2A middle histogram more gentle sub-picture is as reference picture, to another
Each pixel value of piece image is adjusted, and the another piece image is adjusted by the illumination of a sub-picture；
Step4.3：Using the two images after adjustment as input picture, and two width are inputted using weighted average fusion method
Image is merged.
Compared with prior art, the invention has the advantages that：
The invention discloses a kind of RANSAC improved methods towards image mosaic, this method is by extracting characteristics of image
Point, images match is carried out using improved SURF algorithm, with improved RANSAC methods by improving the process of stochastical sampling, is eliminated
Error matching points, improve matching primitives speed and matching accuracy rate, shorten the algorithm time.In the image co-registration stage first to defeated
Enter image and carry out luminance proportion pretreatment, then merged using weighted mean method, so as to preferably complete image mosaic.
Brief description of the drawings
Fig. 1 is a kind of flow chart of RANSAC improved methods towards image mosaic in the present invention.
Fig. 2 is improvement RANSAC algorithm flow charts in the present invention.
Fig. 3 is RANSAC straight line estimation figures in the present invention.
Fig. 4 is image irradiation cumulative distribution function figure in a balanced way in the present invention.
Fig. 5 a, Fig. 5 b are that the improved RANSAC methods before processing based on image mosaic is passed through in the specific embodiment of the invention
Artwork.
Fig. 5 c are the effects after the improved RANSAC methods processing based on image mosaic in the specific embodiment of the invention
Figure.
Embodiment
Below so that common two images are spliced as an example, with reference to accompanying drawing to proposed by the present invention a kind of towards image mosaic
RANSAC improved methods be described in further details.
Images match is searching in two images that are being gathered in different images equipment and there is certain lap
To the point coincided, then carry out spatial alternation to complete the process of image alignment operation using the point coincided searched out.
The difference of information is utilized according to matching, image matching method can substantially be summarized as following 3 class：Based on gray scale, feature based and
Image matching method based on transform domain etc..
The present invention by SURF (Speeded-Up Robust Features) algorithm thought, based on Sobel operators,
When rejecting is mismatched, using improved probability random sampling uniformity (RANSAC) algorithm proposed by the present invention, by constantly adjusting
Whole sampled probability, so as to obtain correct model with less time, improves efficiency of algorithm.In image co-registration, brightness is first carried out
Equilibrium pretreatment, then merged with weighting smoothing algorithm, so as to eliminate the splicing line and intermediate zone of result figure, improve image and spell
Connect quality.
As shown in Figure 1, the RANSAC improved method detailed processes of the invention towards image mosaic are as follows：
The first step, feature point extraction and the generation of description.
Sobel operators are mainly by centered on pixel (i, j) of image f (i, j), in certain limit around
Pixel value do weighting processing, and then judge whether the pixel is in extreme value state, Sobel operators are by extreme value state
Pixel is labeled as marginal point.
Step1.1：Feature point extraction.
Feature point extraction detailed process is as follows：
(1) Sobel operators carry out convolutional calculation using 3 × 3 module, and its Jacobian matrix is
(2) useWithTwo convolution masks, utilize two convolution
Template calculates transverse edge detection G respectivelyxG is detected with longitudinal edgey：
(3) by Gx, GyIt can calculate
Then the threshold value set in advance and G (x, y) are compared, if threshold value is less than G (x, y), then the pixel (i, j) is
Marginal point, characteristic point is extracted as by the marginal point.
Step1.2：The generation of son is described.
After the complete extreme point of image detection, next the extreme point detected is calculated using improved SURF proposed by the present invention
The generation of son is described in method, is specifically divided into following steps progress：
(2.1) centered on the characteristic point that Step1.2 is tried to achieve, pair radius is any point in 6 σ circle domain, asks the circle domain
The haar small echo response coefficients of interior any point in the x and y direction.
(2.2) distance of any point in the round domain to feature dot center is a radius, then with a radius
For starting point, the sector region for taking 50 ° is zoning, and it is square counterclockwise in 6 σ circle domain in the radius then to make the sector region
To rotation, to x and y directions haar wavelet coefficients in this sector region and calculated respectively every 30 °.
(2.3) the haar wavelet coefficients and form a vector, a total of 12 vectors determine mould in this 12 vectors
A maximum vector, then using a maximum vectorial direction of the mould as characteristic point principal direction.
(2.4) and then a characteristic point detected by Step1 is chosen, and centered on this feature point, coordinate system is revolved
The principal direction is gone to, and makees the square area that a length of side is 20 σ, the square area is divided into 4 × 4 zonule,
The haar wavelet coefficients in the x and y directions of each zonule pixel are calculated respectively, and the length of side of each zonule when calculating is
The σ of 5 σ × 5, direction is consistent with principal direction.
(2.5) the haar wavelet systems in the x and y directions of each pixel in the square area for trying to achieve step (2.4)
Number carries out sum operation, respectively obtains dxAnd dy, then by the exhausted of the haar wavelet coefficients of the every bit pixel of the square area
To value summation, ∑ d is respectively obtainedx, ∑ dy。
(2.6) a vector is obtained in each zonule：
V=[∑ dx,∑dy,∑|dx|,∑|dy|],
A total of 16 zonules in the square, the window vector of all zonules constitutes description, its
Dimension size is 4 × 4 × 4=64.
Second step, characteristic matching.
In the matching process, measuring similarity of the present invention uses Euclidean distance, i.e., with minimum distance with time closely
The ratio between, if this than value be less than some given threshold, then it is assumed that be best match, otherwise abandon；The threshold value sets bigger, then
The number of pairs arrived is more, otherwise, and an obtained number of pairs is fewer, the similarity judgment threshold value used in the present invention
For 0.8.
3rd step, image registration is carried out using improved RANSAC methods.
According to the matching characteristic point pair found, it can calculate between the coordinate transformation relation between image, i.e. two images
Transformation matrix.
Projective transformation is transformed between definition two images, projective transformation puts to the proof H and is：
If p=(x, y), q=(x', y') are the characteristic points pair of matching, then projective transformation formula is：
H each free degree parameter h can be calculated by projective transformation formulai(i=0,1,2 ..., 7), and with each free degree
Parameter refines H by iteration, can further determine that the correspondence of characteristic point as initial value, untill corresponding points invariable number,
Stop iteration, so as to complete image rough registration.
RANSAC is a kind of mathematical modeling that parameter is estimated using iterative method, and many methods are using RANSAC to image
Transformation matrix is refined, and is mismatched pair with rejecting, and improves image registration accuracy.Rejected using RANSAC when mismatching, it is traditional
RANSAC appoints first in all match points takes determine that straight line carries out fitting a straight line at 2 points, in error of fitting distance range
Point be referred to as interior point, be otherwise exterior point, RANSAC straight lines estimate as shown in Figure 3.
Then, point is concentrated and recalculates a new straight line and be fitted inside, by continuous iteration, until finding one
The fitting a straight line of individual most imperial palace points amount is used as fitting result, you can rejecting is mismatched a little.
And improved RANSAC algorithms of the invention utilize the minimum increment extracted every time by improving the process of stochastical sampling
Original estimation model parameter, the probability being selected during subsample is sampled below is adjusted further according to the model parameter.By continuous
Iteration is extracted after inspection, has gradually stepped up the probability of the correct temporary pattern of acquisition, so that it is correct to take less time acquisition
Model, is effectively improved the efficiency of algorithm.
If data point sum during image registration is n, data point sample weight isJust
Beginning smallest subset is ψ0.When estimating model parameter, datum error is ei(i=1,2 ..., n)；Sampled probability is：
Probability threshold value is p；Point set S, (S in interimi,Sj) be arbitrfary point pair in S, its Euclidean distance d=| | Si-Sj| |,
Euclidean distance threshold value is D.
Improve RANSAC algorithm flows as shown in Figure 2, it is specific as follows：
Step3.1：Initialization data point sampling weight isSelect initial smallest subset
ψ0；
Step3.2：Computation model parameter and error ei(i=1,2 ..., n)；
Step3.3：Update sample weights ωi, obtain sampled probability p (xi)；
Step3.4：According to sampled probability p (xi), obtain new data subset ψ0；
Step3.5：If p (xi) ＞ P, then Step3.1 is returned, otherwise, that is, temporary pattern is found and obtains point set in interim
S；
Step3.6：If D ＜ d, the final interior point set and final model parameter for meeting model are obtained, terminates algorithm；If d
＞ D, then return to Step3.5 and continue iteration, chooses in interim in point set S arbitrfary point to (Si,Sj), until D ＜ d, acquisition meets mould
The final interior point set of type and final model parameter, terminate algorithm.
4th step, improved image co-registration.
In order to obtain natural splicing result figure, the present invention carries out brightness using histogram equalization method to input picture
Pretreatment, makes input image lightness reach equilibrium, then reuses weighted average fusion method and carries out image co-registration, so both will not
Intermediate zone is produced, splicing line can be effectively eliminated again.It is specific as follows：
Step4.1：Overlay region image histogram normalization is carried out, to obtain the probability density function of overlay region pixel value
Equalize formula.Overlay region image histogram calculating is carried out first, because overlay region image histogram represents overlay region pixel value
Frequency, so the total pixel of each of which frequency divided by overlay region can be obtained into the Probability p (ω) that each pixel value occurs, thus
Cumulative distribution function to overlay region pixel value is as follows：
In above formula, ω is aleatory variable.
Step4.2：According to formula (9), input image pixels value mapping relations are set up, each pixel to image is reflected
Penetrate adjustment.
If input picture is I1、I2, take histogram more gentle image (such as I1) as reference picture, to I2It is each
Pixel value (0≤g≤255) is adjusted, and makes I2By I1Illumination adjustment, as shown in Figure 4.
From accompanying drawing 4, reference picture and image to be adjusted are respectively in g1And g2There is identical cumulative distribution under pixel value
Functional value, i.e.,：FCD(g1)=FCD(g2).Then, I2All pixels value is g in image1Pixel-map be g2.According to the above
Similar operations, to image I2All gray levels be adjusted.Assuming that in a certain gray level, if image to be adjusted and with reference to figure
As having identical cumulative distribution function value, then image pixel value to be adjusted is adjusted to the corresponding pixel value of reference picture；Such as
Fruit reference picture and image to be adjusted do not have identical cumulative distribution function value, then image to be adjusted is kept in a certain gray level
Original pixel value.
Step4.3：Using the two images after adjustment as input picture, and two width are inputted using weighted average fusion method
Image is merged.
The method proposed in the present invention can actually be embedded in FPGA realizations, and exploitation has the camera of image mosaic function or taken the photograph
Camera.Above example only plays a part of explaining technical solution of the present invention, and protection domain of the presently claimed invention does not limit to
In realizing system and specific implementation step described in above-described embodiment.Therefore, only to specific formula and calculation in above-described embodiment
Method is simply replaced, but its substantive content still technical scheme consistent with the method for the invention, all should belong to the present invention
Protection domain.
Claims (7)
1. a kind of RANSAC improved methods towards image mosaic, it is characterized in that, comprise the following steps：
The first step：Feature point extraction and the generation of description；
Second step：Characteristic matching；
3rd step：Image registration is carried out using improved RANSAC methods；
4th step：Brightness pretreatment is carried out to input picture using histogram equalization method, input image lightness is reached
Weighing apparatus, then reuses weighted average fusion method and carries out image co-registration.
2. a kind of RANSAC improved methods towards image mosaic according to claim 1, it is characterised in that described first
Feature point extraction specifically includes following steps described in step：
(1.1) Sobel operators carry out convolutional calculation using 3 × 3 module, and its Jacobian matrix is
(1.2) useWithTwo convolution masks, utilize two convolution moulds
Plate calculates transverse edge detection G respectivelyxG is detected with longitudinal edgey：
(1.3) by Gx, GyIt can calculate
Then the threshold value set in advance and G (x, y) are compared, if threshold value is less than G (x, y), then the pixel (i, j) is edge
Point, characteristic point is extracted as by the marginal point.
3. a kind of RANSAC improved methods towards image mosaic according to claim 2, it is characterised in that described first
The generation that son is described described in step comprises the following steps：
(2.1) centered on the characteristic point of extraction, pair radius is any point in 6 σ circle domain, asks any point in the circle domain to exist
Haar small echo response coefficients on x and y directions；
(2.2) any point in the round domain to the distance of feature dot center be radius, then using a radius for
Point, the sector region for taking 50 ° is zoning, then makes the sector region counterclockwise be revolved in the circle domain that the radius is 6 σ
Turn, to x and y directions haar wavelet coefficients in this sector region and calculated respectively every 30 °；
(2.3) the haar wavelet coefficients and form a vector, a total of 12 vectors determine in this 12 vectors that mould is maximum
A vector, then using a maximum vectorial direction of the mould as characteristic point principal direction；
(2.4) and then a characteristic point detected by Step1 is chosen, and centered on this feature point, coordinate system is rotated to
The principal direction, and make the square area that a length of side is 20 σ, the square area is divided into 4 × 4 zonule, respectively
Calculate the haar wavelet coefficients in the x and y directions of each zonule pixel, the length of side of each zonule when calculating for 5 σ ×
5 σ, direction is consistent with principal direction；
(2.5) the haar wavelet coefficients in the x and y directions of each pixel in the square area for trying to achieve step (4) are carried out
Sum operation, respectively obtains dxAnd dy, then the absolute value of the haar wavelet coefficients of the every bit pixel of the square area is asked
With respectively obtain ∑ dx, ∑ dy；
(2.6) a vector v=[∑ d is obtained in each zonulex,∑dy,∑|dx|,∑|dy|], the square area bag
16 zonules are included, the window vector of 16 zonules constitutes description, and its dimension size is 4 × 4 × 4=64.
4. a kind of RANSAC improved methods towards image mosaic according to claim 1, it is characterised in that described second
In step, in the matching process, measuring similarity is carried out using Euclidean distance, i.e., with minimum distance with the ratio between time closely, if should
Ratio is less than some given threshold, then it is assumed that is best match, otherwise abandons；
。
5. a kind of RANSAC improved methods towards image mosaic according to claim 1, it is characterised in that the described 3rd
In step, by improving the process of stochastical sampling, model parameter is estimated using the minimum subsample extracted every time, further according to estimation
Model parameter adjustment subsample sample below in the probability that is selected, and be constantly iterated extraction inspection to improve acquisition
The probability of correct temporary pattern.
6. a kind of RANSAC improved methods towards image mosaic according to claim 4, it is characterised in that the described 3rd
In step, data point sum when defining image registration is n, and data point sample weight isInitially most
Small subset is ψ0；
When estimating model parameter, datum error is ei(i=1,2 ..., n), sampled probability is：
P probability threshold values are, it is interim in point set S, (Si,Sj) be arbitrfary point pair in S, its Euclidean distance d=| | Si-Sj| |, D Euclideans away from
From threshold value；
Improve RANSAC algorithm flows specific as follows：
Step3.1：Initialization data point sampling weight isSelect initial smallest subset ψ0；
Step3.2：Computation model parameter and error ei(i=1,2 ..., n)；
Step3.3：Update sample weights ωi, obtain sampled probability p (xi)；
Step3.4：According to sampled probability p (xi), obtain new data subset ψ0；
Step3.5：If p (xi) ＞ P, then Step3.1 is returned, otherwise, that is, temporary pattern is found and obtains point set S in interim；
Step3.6：If D ＜ d, the final interior point set and final model parameter for meeting model are obtained, terminates algorithm；If d ＞ D,
Then return to Step3.5 and continue iteration, choose in interim in point set S arbitrfary point to (Si,Sj), until D ＜ d, acquisition meets model
Point set and final model parameter in final, terminate algorithm.
7. a kind of RANSAC improved methods towards image mosaic according to claim 1, it is characterised in that the described 4th
Step specifically includes following steps：
Step4.1：Overlay region image histogram is normalized, the equilibrium of the probability density function of overlay region image pixel value is obtained
Change formula, the cumulative distribution function of overlay region pixel value is as follows：
In formula, ω is aleatory variable, and p (ω) is the probability that each pixel value occurs；
Step4.2：According to formula (7), input image pixels value mapping relations are set up, each pixel to image carries out mapping tune
It is whole；Definition input picture is I1、I2, take I1、I2A middle histogram more gentle sub-picture is as reference picture, to another width
Each pixel value of image is adjusted, and the another piece image is adjusted by the illumination of a sub-picture；
Step4.3：Using the two images after adjustment as input picture, and using weighted average fusion method to two width input pictures
Merged.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710236500.9A CN107103579A (en) | 2017-04-12 | 2017-04-12 | A kind of RANSAC improved methods towards image mosaic |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710236500.9A CN107103579A (en) | 2017-04-12 | 2017-04-12 | A kind of RANSAC improved methods towards image mosaic |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107103579A true CN107103579A (en) | 2017-08-29 |
Family
ID=59675222
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710236500.9A Pending CN107103579A (en) | 2017-04-12 | 2017-04-12 | A kind of RANSAC improved methods towards image mosaic |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107103579A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109523495A (en) * | 2018-10-15 | 2019-03-26 | 北京东软医疗设备有限公司 | Image processing method and device, equipment and storage medium |
| CN110852978A (en) * | 2019-08-28 | 2020-02-28 | 北京联合大学 | Outlier removing method used before saliency map fusion |
| CN111127318A (en) * | 2019-12-05 | 2020-05-08 | 南京莱斯电子设备有限公司 | Panoramic image splicing method in airport environment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20120021666A (en) * | 2010-08-12 | 2012-03-09 | 금오공과대학교 산학협력단 | Panorama image generating method |
| CN105608671A (en) * | 2015-12-30 | 2016-05-25 | 哈尔滨工业大学 | Image connection method based on SURF algorithm |
-
2017
- 2017-04-12 CN CN201710236500.9A patent/CN107103579A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20120021666A (en) * | 2010-08-12 | 2012-03-09 | 금오공과대학교 산학협력단 | Panorama image generating method |
| CN105608671A (en) * | 2015-12-30 | 2016-05-25 | 哈尔滨工业大学 | Image connection method based on SURF algorithm |
Non-Patent Citations (2)
| Title |
|---|
| 李海洋等: "一种改进的基于SURF特征匹配的图像拼接算法", 《太原理工大学学报》 * |
| 王小平等: "基于Sobel算子和改进SURF算法的图像拼接方法", 《电视技术》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109523495A (en) * | 2018-10-15 | 2019-03-26 | 北京东软医疗设备有限公司 | Image processing method and device, equipment and storage medium |
| CN110852978A (en) * | 2019-08-28 | 2020-02-28 | 北京联合大学 | Outlier removing method used before saliency map fusion |
| CN111127318A (en) * | 2019-12-05 | 2020-05-08 | 南京莱斯电子设备有限公司 | Panoramic image splicing method in airport environment |
| CN111127318B (en) * | 2019-12-05 | 2023-04-07 | 南京莱斯电子设备有限公司 | Panoramic image splicing method in airport environment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107301661B (en) | High-resolution remote sensing image registration method based on edge point features | |
| Park et al. | Look wider to match image patches with convolutional neural networks | |
| CN109767463B (en) | Automatic registration method for three-dimensional point cloud | |
| CN104318548B (en) | Rapid image registration implementation method based on space sparsity and SIFT feature extraction | |
| CN108108746B (en) | License plate character recognition method based on Caffe deep learning framework | |
| CN108898610A (en) | A kind of object contour extraction method based on mask-RCNN | |
| CN108364272A (en) | A kind of high-performance Infrared-Visible fusion detection method | |
| CN109584282B (en) | Non-rigid image registration method based on SIFT (scale invariant feature transform) features and optical flow model | |
| CN109147040B (en) | Template-based human point cloud hole repairing method | |
| CN111126412B (en) | Image key point detection method based on characteristic pyramid network | |
| CN104809731B (en) | A kind of rotation Scale invariant scene matching method based on gradient binaryzation | |
| CN108765476B (en) | Polarized image registration method | |
| CN108550166B (en) | Spatial target image matching method | |
| CN112819772A (en) | High-precision rapid pattern detection and identification method | |
| CN110135438B (en) | Improved SURF algorithm based on gradient amplitude precomputation | |
| CN105787943B (en) | SAR image registration method based on multi-scale image block feature and rarefaction representation | |
| CN108428249A (en) | A kind of initial position and orientation estimation method based on optical flow tracking and double geometrical models | |
| CN106534833B (en) | A kind of double vision point three-dimensional video-frequency antihunt means of joint space time shaft | |
| CN111709980A (en) | Multi-scale image registration method and device based on deep learning | |
| CN113888461A (en) | Method, system and equipment for detecting defects of hardware parts based on deep learning | |
| CN107103579A (en) | A kind of RANSAC improved methods towards image mosaic | |
| CN103080979A (en) | System and method for synthesizing portrait sketch from photo | |
| CN111192194A (en) | Panoramic image splicing method for curtain wall building vertical face | |
| CN111783548B (en) | SAR image and visible light image matching method based on improved feature extraction and game theory hypergraph | |
| CN111009005A (en) | Scene classification point cloud rough registration method combining geometric information and photometric information |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170829 |
|
| RJ01 | Rejection of invention patent application after publication |