CN101840570A - Fast image splicing method - Google Patents

Abstract

The invention discloses a fast image splicing method, which is characterized by comprising the following steps of: (1) acquiring a single-breadth field image; (2) extracting characteristic information in an image to be spliced; (3) performing two-dimensional wavelet transform on the image to be spliced to obtain high-frequency coefficients reflecting the profile of change in a horizontal direction and a vertical direction of the image, and realizing overlapped position alignment in combination with a phase correlation method; and (4) adopting a weighted mean method to fuse overlapped areas of the image to finish the splicing. In the method, low-energy coefficients reflecting the gradual change of illumination are removed by performing the two-dimensional wavelet transform on the image and then the high-frequency coefficients reflecting the profile of the change in the horizontal direction and the vertical direction of the image are obtained, which can effectively reduce calculation complexity; and simultaneously, the fast and accurate overlapped position alignment is realized in combination with the phase correlation method and the fusion of the image is realized by adopting the weighted mean method so as to realize the fast and accurate splicing of PCB images of various specifications and various colors.

Description

A kind of fast image splicing method

Technical field

The invention belongs to graphic processing method, relate in particular to a kind of fast image splicing method based on two-dimensional wavelet transformation.Can be applicable to mount the image mosaic in the PCB automatic optical detecting system.

Background technology

Along with science and technology development, image mosaic has become very important technology in the Image Information Processing field.The image mosaic technology is according to doubling of the image part, and the image of many linkings is pieced together a high-resolution panorama sketch.The image mosaic technology is widely used in fields such as digital video, motion analysis, virtual reality technology, medical image analysis, remote sensing image processing at present, directly studies the few of the middle image mosaic problem of AOI (automated optical detection).Image mosaic comprises the image registration of lap and image co-registration two parts content of seam crossing.The core of image mosaic technology and key are the image registration problems, and it is directly connected to the success ratio and the travelling speed of merging algorithm for images.Therefore, the research of registration Algorithm is the emphasis of Study of Image Mosaics Technology for many years.As far back as 1992, the Lisa Gottesfeld Brown of Cambridge University has just summed up the basic theories of each field image registration techniques and their main method, and these fields include medical image analysis, Remote Sensing Data Processing and computer vision, pattern-recognition etc.1996, Richard Szeliski proposed based drive panoramic picture split-join model, adopted the non-linear Method for minimization of Levenberg-Marquardt iteration (being called for short the L-M algorithm), concerned and carried out image registration by obtaining geometric transformation between image.Because the method effect is better, fast convergence rate, and can handle have translation, the image to be spliced of rotation, multiple conversion such as affine, therefore also become the classic algorithm in image mosaic field, and from then on Richard Szeliski also becomes the founder in image mosaic field, the theory that he proposed has become a kind of classical theory system, a lot of people is still arranged in his splicing theory of research up to today.2000, Shmuel Peleg, BennyRousso, AlexRav-Acha and Assaf Zomet have done further improvement on the basis of Richard Szeliski, adaptive image mosaic model has been proposed, it is the different motion according to camera, adaptively selected split-join model carries out the splicing that image is finished in multiple projection by image being divided into slat shape.This achievement in research has promoted further developing of image mosaic technology undoubtedly, and from then on adaptive problem also becomes the new focus of image mosaic area research.

Meanwhile, the development of other registration Algorithm also has been applied on the image mosaic technology.Except above-mentioned classic algorithm, also have two kinds of main method, a kind of is phase correlation method, another kind is based on the method for registering images of geometric properties.Phase correlation method was proposed by Kuglin and Hines as far back as 1975, had the scene independence, the image of pure two-dimension translational accurately can be alignd.Afterwards, Decastro and Morandi found can make the rotation of image be converted into the translation of image with Fourier transform, and this method is very suitable to the image registration of rotation and convergent-divergent.1996, Reddy and Chaterji improved the algorithm of Decastro, had significantly reduced the quantity that needs conversion.The translation vector of two width of cloth images can directly calculate by the phase place of their cross-power spectrums (CrossPower Spectrum).The registration that the application Fourier transform carries out image is another achievement in research in image mosaic field, and along with use the maturation of Fourier transform in the proposition and the signal Processing field of fast fourier transformation algorithm, the image mosaic technology has also obtained corresponding development.

Method for registering images based on geometric properties is another research focus of image mosaic technology.1994, Blaszka T and Rachid Deriche filtered by two-dimentional Gaussian Blur and obtain some low-level features models, as limit model, angle model and summit model.Based on this, there was increasing people to begin one's study afterwards and carried out the method for image mosaic based on the low-level features in these images.1997, Zoghlami I, FaugerasoDeriche R proposed the image alignment algorithm based on how much angle models.Then in 1999, the vision mode that Bao P and Xu D propose to utilize wavelet transformation to extract reservation limit (edge-preserving) carried out image alignment, and Nielsen F then proposes the matching process based on the geometric point characteristic optimization.2000, Kang.E.Cohen I and Medioni G proposed to carry out based on the advanced features of image the method for image mosaic, and he utilizes the characteristic image graph of a relation to carry out image alignment.Utilize its advanced features finally by the low-level features of utilizing image, people are deep day by day to the analysis and the understanding of image, and the image mosaic Study on Technology is also grown up gradually.

According to registration according to different, method for registering images roughly can be divided into based on the image registration algorithm of feature with based on the image registration algorithm in zone.The degree that image registration algorithm participates in according to the people can be divided into semi-automatic registration and full-automatic registration.In addition, to carry out image synthetic owing to directly get in two width of cloth images wherein the width of cloth in the overlapping region, and after the image registration, the boundary that is easy in the overlapping region forms tangible seam, and these all can make follow-up image segmentation etc. handle and produce error.For this class problem, all be to adopt the method for image co-registration to solve usually.And image interfusion method commonly used has the method for being fade-in fade-out, turriform blending algorithm, mathematics form method, wavelet algorithm etc.

Summary of the invention

It is bad to handle the illumination interference effect at traditional merging algorithm for images, and the algorithm complexity, is not suitable for being applied in the problem of the image mosaic among the AOI.The invention provides a kind of fast image splicing method that is applicable to AOI that the time is shorter, accuracy is higher that splices based on two-dimensional wavelet transformation.

The present invention includes following technical characterictic: a kind of fast image splicing method is characterized in that comprising the steps:

(1) obtains the single width field-of-view image;

(2) characteristic information among the extraction figure to be spliced;

(3) treat stitching image and carry out two-dimensional wavelet transformation, obtain the high frequency coefficient that reflection image level and vertical direction change profile, realize the lap position registration in conjunction with phase correlation method then;

(4) adopt method of weighted mean to come doubling of the image zone is merged, finish splicing.

Further, described step (3) is specific as follows:

(31) image after the gray scale processing is carried out two-dimensional wavelet transformation, reduce the resolution of image, compressed image; Obtain the high frequency coefficient that reflection image level direction and vertical direction change simultaneously;

Ignore image vertical direction deviation when (32) level is spliced, ignore the deviation of image level direction when vertically splicing; Adopt the gray scale column mean of vertical high frequency coefficient during the level splicing, the horizontal high frequency coefficient gray scale of employing row average is carried out computing when vertically splicing;

(33) according to concrete image setting threshold value, eliminate because uneven illumination is spared the noise that causes, improve the precision of registration;

(34) according to the registration position on phase correlation method acquisition image level and the vertical direction.

Further, described step (4) seamlessly transits the registration position that step (34) obtains for using method of weighted mean, eliminates the vestige of splicing, realizes image co-registration.

The present invention compared with prior art, be primarily aimed at the engineering reality of AOI, remove the low energy coefficients that reflects the illumination gradual change by image being carried out two-dimensional wavelet transformation, obtain the high frequency coefficient that reflection image level and vertical direction change profile, can reduce computation complexity effectively; Realize fast accurate lap position registration and adopt method of weighted mean to realize the fusion of image in conjunction with phase correlation method simultaneously.To realize that the PCB image of plurality of specifications, different colours is realized splicing fast and accurately.

Description of drawings

Fig. 1 is an image mosaic process flow diagram of the present invention;

Fig. 2 is the high frequency coefficient after image wavelet decomposes;

Fig. 3 is the figure as a result of image level phase place related operation;

Fig. 4 is the horizontal splicing effect figure that directly gets two width of cloth gradation of image values;

Fig. 5 is the horizontal splicing effect figure that adopts method of weighted mean to merge.

Embodiment

The present invention is a kind of fast image splicing method, as shown in Figure 1, comprises the steps:

(1) obtains the single width field-of-view image;

(2) characteristic information among the extraction figure to be spliced;

(3) treat stitching image and carry out two-dimensional wavelet transformation, obtain the high frequency coefficient that reflection image level and vertical direction change profile, realize the lap position registration in conjunction with phase correlation method then;

(4) adopt method of weighted mean to come doubling of the image zone is merged, finish splicing.

Wherein key is the fusion for the registration and the image of image, and following mask body is illustrated these two aspects:

One, image registration:

Image registration is based on the profile of image, and the high frequency coefficient of wavelet decomposition has just in time reflected the variation profile of image.In the engineering reality of paster install machinery vision-based detection AOI, the paster production line has the real-time requirement to the detection speed of AOI.And the operand of phase correlation method is very big, establishes two width of cloth image f ₁(x, y) and f ₂Have coordinate transformation relation as follows between (x ', y '):

$\left[\begin{array}{c}{x}^{\′}\\ y^{\′}\end{array}\right]=\left[\begin{array}{cc}\cos {\mathrm{\θ}}_0& \sin {\mathrm{\θ}}_0\\ -\sin {\mathrm{\θ}}_0& \cos {\mathrm{\θ}}_0\end{array}\right]\left[\begin{array}{c}x-x_0\\ {y-y}_0\end{array}\right]---\left(1\right)$

θ wherein ₀The expression rotation amount, (x ₀, y ₀) the expression translational movement.Both all get as Fourier transform:

$F_2(\mathrm{\ζ},\mathrm{\η})=F_1(\mathrm{\ζ}\cos {\mathrm{\θ}}_0+\mathrm{\η}\sin {\mathrm{\θ}}_0,-\mathrm{\ζ}\sin {\mathrm{\θ}}_0+---\left(2\right)$

$\mathrm{\ηos}{\mathrm{\θ}}_0)e^{-j[(\mathrm{\ζ}\cos {\mathrm{\θ}}_0+\mathrm{\η}\sin {\mathrm{\θ}}_0)x_0+(-\mathrm{\ζ}\sin {\mathrm{\θ}}_0+\mathrm{\ηos}{\mathrm{\θ}}_0)y_0]}$

As seen, the translation in the spatial domain only is reflected in phase change in frequency domain, and the rotation amount in the spatial domain can have the amplitude of frequency domain to determine.The cross-power spectrum of two width of cloth images is:

$\frac{F_1(\mathrm{\&zeta;},\mathrm{\&eta;}){{\mathrm{<figure-callout\; id="2"\; label="F"\; filenames="HSA00000099537200013.png"\; state="\{\{state\}\}">F</figure-callout>}}_2}^{*}(\mathrm{\&zeta;},\mathrm{\&eta;})}{\left|F_1(\mathrm{\&zeta;},\mathrm{\&eta;}){F_2}^{*}(\mathrm{\&zeta;},\mathrm{\&eta;})\right|}=e^{-j(\mathrm{\&zeta;}{x}_0+\mathrm{\&eta;}{y}_0)}---\left(3\right)$

The template of M * N is (M * N) by the calculated amount of formula (3) ², can't satisfy the requirement of detection speed like this.And the image that splicing obtains is just in order to do path planning and location etc., less demanding on the pixel.Can carry out two-dimensional wavelet transformation to image, the image of such width of cloth 1200dpi resolution reduces resolution to 1/16 after the conversion, i.e. 80 μ m/pixel, size for 0201 element 0.3mm of minimum is 3 pixels on image, still can differentiate, and can satisfy follow-up demand.

Therefore image is carried out two-dimensional wavelet transformation, obtain low-frequency component respectively and reflect image level variation, vertical change respectively and the high frequency coefficient of angular direction variation totally 4 parts.Figure 2 shows that the result of two-dimentional multiple dimensioned decomposition, from left to right is level after the wavelet decomposition, vertical and to the high frequency coefficient on the angular direction successively.That wherein adopt is the wave filter bior3.7 of Bi-orthogonal Spline Wavelet Transformation correspondence, exponent number N _rEqual 3, N _dEqual 7, change of scale is 2.

According to the character of the grey scale change of horizontal direction in the wavelet decomposition Vertical factor of the image reflection image, i.e. the vertically profiling of object in the image, and the vertically profiling foundation of horizontal departure registration between image just.Therefore, when handling the splicing of image level direction, adopt the high frequency coefficient on the vertical direction of image wavelet decomposition back.When image wavelet transform obtains the high frequency coefficient of image, also image has been realized compression.The image of one 1200 * 910 pixel, 1092000 bytes, the image of boil down to 311 * 238 pixels, 592144 bytes is compressed to 54% of original image after conversion.The calculated amount that the back phase place is relevant has just significantly reduced like this.

The X-Y platform mobile accuracy of AOI is quite high, and error is at 6-10 μ m, and in the level splicing, we can ignore the deviation on the vertical direction like this; In vertical splicing, ignore the deviation on the horizontal direction.We can consider directly to adopt column mean in the level splicing, directly adopt the row average in vertical splicing.The column mean of one width of cloth M * N image and row average are respectively:

$X\left(j\right)=\underset{i}{\overset{M}{\mathrm{\&Sigma;}}}F(i,j)/M---\left(4\right)$

$Y\left(i\right)=\underset{j}{\overset{N}{\mathrm{\&Sigma;}}}F(i,j)/N---\left(5\right)$

Wherein F (i, j) be image (i, the pixel value of j) locating, M are picturedeep,

N is a picturewide.X (i) and Y (j) are respectively the gray scale column mean and the capable gray scale row averages of i of j row.

In AOI, because the cost of lighting source and light source be along with the variation of service time, ubiquity the even phenomenon of uneven illumination.The image of camera acquisition is subjected to the illumination interference of noise easily like this, influences the result of registration.These noises are cashed at high-frequency domain and are low-energy coefficient.Except only adopting the high frequency coefficient of reflection profile variations, also, absolute value is fallen less than the noise filtering of regarding as of threshold value when we do registration for this situation by setting threshold in the wavelet decomposition high frequency coefficient.Wherein the setting of threshold value does not have strict regulation, and we are that the high frequency coefficient maximum amplitude according to concrete image multiply by a scale-up factor and determines.

With Matlab as experiment porch, with a among Fig. 1 figure as with reference to image, b figure is an image to be spliced, vertical high frequency coefficient behind two width of cloth image wavelet transforms is carried out Fast Fourier Transform (FFT), by formula (1) carries out the phase place correlation computations, obtain on the image level direction correlated results as shown in Figure 3.As seen from the figure, obtained peak effect preferably, algorithm accurately finds the lap position at 172 pixel places of reference picture horizontal direction.

Two, based on the image co-registration of method of weighted mean

After finding the lap position of two width of cloth images,,, tangible seam will occur, influence the effect of image in registration position because two width of cloth gradation of image values is discontinuous if directly splice.As shown in Figure 4.

Consider our X-Y mobile platform precision height, the rotation error between image is very little, and just has the translational movement of a level or a direction of vertical centering control between image, and we have adopted method of weighted mean to eliminate the seam of splicing place.Method of weighted mean is a kind of simple Image Fusion.The principle of this method is exactly that two width of cloth image corresponding pixel points be multiply by behind the weighting coefficient image that addition again obtains merging.If image I 1 (x, weighting coefficient y) are α, and then fused images can be expressed as:

I(i，j)＝αI ₁(i，j)+(1-α)I ₂(i，j) (6)

Wherein, weighting coefficient α satisfies 0≤α≤1.If α=0.5 then is equivalent to two width of cloth images and averages; If it is α is taken as the gradual change coefficient, promptly also different in different coordinate points values; When α slowly changed to 0 by 1, image was from I ₁(i j) slowly carries out the transition to I ₂(i j), so just can realize seamlessly transitting of image, thereby has eliminated the vestige of splicing, reaches the target of image co-registration.Figure 5 shows that the splicing effect that adopts method of weighted mean to merge.As seen algorithm has well been eliminated spliced seam, realizes seamlessly transitting.

Claims (3)

1. a fast image splicing method is characterized in that comprising the steps:

(1) obtains the single width field-of-view image;

(2) characteristic information among the extraction figure to be spliced;

(3) treat stitching image and carry out two-dimensional wavelet transformation, obtain the high frequency coefficient that reflection image level and vertical direction change profile, realize the lap position registration in conjunction with phase correlation method then;

(4) adopt method of weighted mean to come doubling of the image zone is merged, finish splicing.

2. fast image splicing method according to claim 1 is characterized in that described step (3) is specific as follows:

(31) image after the gray scale processing is carried out two-dimensional wavelet transformation, reduce the resolution of image, compressed image; Obtain the high frequency coefficient that reflection image level direction and vertical direction change simultaneously;

Ignore image vertical direction deviation when (32) level is spliced, ignore the deviation of image level direction when vertically splicing; Adopt the gray scale column mean of vertical high frequency coefficient during the level splicing, the horizontal high frequency coefficient gray scale of employing row average is carried out computing when vertically splicing;

(33) according to concrete image setting threshold value, eliminate because uneven illumination is spared the noise that causes, improve the precision of registration;

(34) according to the registration position on phase correlation method acquisition image level and the vertical direction.

3. fast image splicing method according to claim 2 is characterized in that described step (4) seamlessly transits the registration position that step (34) obtains for using method of weighted mean, eliminates the vestige of splicing, realizes image co-registration.

Images