CN107301620A - Method for panoramic imaging based on camera array - Google Patents

Abstract

The invention discloses a panoramic imaging method based on a camera array, which mainly solves the problems of small stitching scenes and "ghost images" in the prior art. The solution is: 1) Use an array camera to acquire multiple images, read in two images to extract SIFT features respectively, perform feature matching search, obtain matching points of the two images and perform screening, calculate the optimal transformation matrix, and according to the optimal transformation The matrix transforms the image, and uses the improved best suture algorithm and weighted average fusion algorithm to fuse and stitch the image; 2) Repeat step 1) until the upper and lower parts of the image stitching are completed, so that it is used as the image to be stitched and rotated counterclockwise Continue stitching after 90°, and rotate the result 90° clockwise to get the final stitched panoramic image. The invention significantly eliminates the "ghost" phenomenon, and obtains a panoramic image with a large field of view and high resolution, which is closer to a real panoramic image, and can be used for splicing larger scene images in horizontal and vertical directions.

Description

Panoramic Imaging Method Based on Camera Array

technical field

The invention belongs to the technical field of image processing, and in particular relates to a panoramic imaging method based on a camera array, which can be used for splicing larger scene images in horizontal and vertical directions.

Background technique

With the development of science and technology, digital imaging technology is gradually stepping up to a new level, and digital imaging equipment has also begun to be widely used in daily life. Using digital cameras, mobile phones and other equipment to take photos has become an indispensable part of people's daily life. At the same time, the limitation of single camera imaging is also becoming increasingly apparent: in some special application scenarios, due to the limitations of digital imaging equipment itself, the needs of users cannot be well met. For example, when people want to obtain wide-field and high-resolution images, they can only use wide-angle cameras in many cases, but their expensive prices make people prohibitive.

In order to solve the above problems, image stitching technology came into being. This technology can match and align a series of small-view images with partially overlapping boundaries according to the corresponding algorithm, and then fuse them, and finally splicing them into a wide-view image.

The most direct application of image stitching technology is the panoramic imaging mode that comes with the mobile phone. However, its limitations are also obvious: it can only be taken horizontally or vertically, and the final image only extends in one direction; at the same time, shooting such an image requires a hand-held There is a very high degree of stability when shooting, otherwise the final image will be deformed, and the desired effect cannot be obtained, which greatly reduces the user experience.

The patent "A method and mobile terminal for panoramic photography" owned by Vivo Mobile Communications Co., Ltd. (application number 201610515352.x, application date 2016.06.30, authorization number CN 1059779156A, authorization date 2016.09.28) proposes a panoramic photography method and mobile terminal. The patented technology includes the first, second and third cameras. By controlling the three cameras in the panoramic shooting to obtain three images, image stitching is performed at the same time to generate a target panoramic image, which can be captured at one time without rotating the mobile terminal horizontally. A panoramic image is obtained by taking a photo. The disadvantage of this method is that only one direction of images can be obtained by taking a photo, and the final result cannot meet the needs of some large scene stitching.

Contents of the invention

The purpose of the present invention is to address the shortcomings of the above-mentioned prior art, and propose a panoramic imaging method based on camera arrays, so as to obtain images in both horizontal and vertical directions at the same time by taking a single photo, so as to meet the needs of larger scene image stitching.

The basic idea of the present invention is: use 2×3 array cameras to obtain array images, and there is an overlapping area between every two images, control six cameras to collect images synchronously; perform feature extraction and pairwise matching on the images, and finally perform image fusion Finished stitching. Its implementation steps are as follows:

(1) Use the array camera to complete the simultaneous acquisition of multiple images, and acquire i images, m≤i;

(2) Read in two images and extract scale-invariant feature transformation feature points, namely SIFT feature points;

(3) carry out feature matching search to step (2) gained SIFT feature point, obtain the matching point of every two images;

(4) screening and calculating optimal transformation matrix H for every two image matching points obtained in step (3);

(5) Transform the image according to the optimal transformation matrix obtained in step (4), and perform image fusion:

(5a) Transform any one of the two input images according to the optimal transformation matrix obtained in step (4), so that the two images are located in the same coordinate system and the two images have overlapping regions;

(5b) Carry out brightness correction to the two images converted to the same coordinate system in (5a), so that the brightness difference of the two images is minimized;

(5c) Find an optimal suture line on the registered image;

(5d) Carrying out weighted average fusion to the rectangle containing the best suture line to obtain the mosaic panoramic image of the two images;

(6) Array image stitching:

(6a) Take the previously stitched image and the image to be stitched as the input two images, continue to repeat steps (2) to (5), and proceed in a loop until the image to be stitched is the mth image, m<= i, finally get the horizontal mosaic panorama of m images;

(6b) Repeat (6a) k times, k<=i and k×m≤i, to obtain k horizontal mosaic images, each horizontal mosaic image is stitched by m images;

(6c) Take the first two of the k horizontal images as input images, rotate them counterclockwise by 90°, repeat steps (2) to (5), and obtain the vertical mosaic of the two images. For the remaining k-2 horizontal images, Take the previously stitched image and the image to be stitched rotated 90° counterclockwise as the input two images, and then repeat steps (2) to (5) in a loop until the image to be stitched is the kth image , and then rotate the stitched longitudinal images clockwise by 90° to finally obtain a horizontal and vertical stitched panorama of i images.

Compared with the prior art, the present invention has the following advantages:

First, the present invention combines the array image with the SIFT algorithm, and improves the SIFT algorithm, so that the splicing time can be greatly reduced when the depth of field is basically the same;

Second, the present invention effectively realizes the combination of the array image and the fusion algorithm, and improves the existing algorithm for finding the best suture line, that is, the array camera is used to acquire images, and i images can be collected simultaneously for splicing, reducing The scene error caused by time changes, especially in the dynamic scene, the object displacement is relatively minimal, which makes the stitching effect better; the improved optimal seam algorithm can effectively avoid moving objects for stitching, and the experimental results show that the stitching image is almost non-existent The frequency of splicing seams and the phenomenon of "ghosting" has also been greatly reduced;

Thirdly, the present invention uses an array camera to collect images, which can be spliced and formed in one shot, reduces workload, and obtains a spliced image with a larger field of view and higher resolution.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Fig. 2 is a camera array diagram for image acquisition in the present invention;

Fig. 3 is six images that camera array gathers among the present invention;

FIG. 4 is a panoramic image after splicing and fusion of the six images in FIG. 3 .

detailed description

Specific embodiments of the present invention are described in detail below in conjunction with accompanying drawing:

With reference to Fig. 1, the realization steps of the present invention are as follows:

Step 1, get the image.

The camera array shown in Figure 2 is used to collect images. The camera array is composed of i cameras in two-dimensional combination in the horizontal and vertical directions. By adjusting the position and focal length of each camera, these cameras can always be arranged in rows and columns and the overall combination is Rectangles to get images that meet different requirements.

Each camera in the camera array can continuously collect multiple images; images of different fields of view can be obtained by changing the position of the camera; at the same time, each camera can adjust the focal length to obtain images with different depths of field.

After adjusting the camera position and focal length, quickly press the start and end keys to obtain several images collected by each camera. The images collected by every two adjacent cameras have overlapping areas, and a total of i images with overlapping areas are obtained. images, this example gives but not limited to six images.

Step 2, read in two images and extract SIFT features respectively.

Common feature point extraction algorithms include: Harris operator, LOG operator, SUSAN operator, SIFT algorithm, etc. The present invention uses a scale-invariant feature transformation algorithm to extract feature points, i.e. SIFT feature points, and the steps are as follows:

(2a) Construct a Gaussian pyramid and a Gaussian difference pyramid, and detect the extreme value of the scale space;

(2a1) The process of constructing a Gaussian pyramid includes two steps of downsampling the image and performing Gaussian smoothing on the image;

According to the original size of the image and the size of the image at the top of the tower, calculate the number of layers n of the pyramid:

n=log2 _{ min(M,N)}-t,t∈[0,log2 _{ min(M,N)})

Among them, M and N are the length and width of the original image respectively, and t is the logarithm value of the minimum dimension of the tower top image.

In the present invention, the focal length of each camera is adjusted so that the depth of field of the images taken by the camera is basically the same, and the scales of the two images are basically the same, so that when the pyramid is constructed for the two images, the number of pyramid layers n is a value greater than 1 and less than 4 , to reduce splicing time;

(2a2) The original image is used as the first layer of the Gaussian pyramid, and then the original image is down-sampled layer by layer. The new image obtained by each down-sampling is a new layer of the pyramid until the nth layer, and a series of From large to small images, form a tower model from bottom to top to get the initial image pyramid;

(2a3) An image of each layer of the initial image pyramid is Gaussian blurred using different parameters, so that each layer of the pyramid contains multiple Gaussian blurred images, and multiple images of each layer of the pyramid are collectively called a group to obtain a Gaussian pyramid;

(2a4) Constructing a Gaussian difference pyramid, i.e. a DOG pyramid: Subtract the adjacent upper and lower images in each group of the Gaussian pyramid obtained in (2a3) to obtain a Gaussian difference pyramid;

(2a5) Carry out spatial extreme point detection:

Take each pixel in each group of the Gaussian difference pyramid, and compare them with all the pixels in the 26 neighborhoods of this picture and the upper and lower two: If the value of the pixel point taken from the Gaussian difference pyramid is the maximum value or the minimum value, the pixel value of the selected point is an extreme value of the scale space of the image at the current scale, where the scale space is realized by a Gaussian pyramid, and each image in each group corresponds to a different scale;

(2b) Use the scale space extremum in (2a) as a key point to locate and determine the direction of the key point:

(2b1) Remove low-contrast points by interpolation and eliminate edge response to complete the precise positioning of key points;

(2b2) Assign direction to feature points:

For (2b1) precisely positioned key points, the gradient and direction distribution features of the pixels in the 3σ neighborhood window of the Gaussian pyramid image where it is located are collected. The modulus m(x, y) and direction θ(x, y) of the gradient are as follows:

θ(x,y)=tan ^-1 ((L(x,y+1)-L(x,y-1))/L(x+1,y)-L(x-1,y)))

Among them, L is the scale space value where the key point is located, the gradient modulus m(x, y) is added according to the Gaussian distribution of σ=1.5σ_oct, and the radius of the 3σ neighborhood window is 3×1.5σ_oct;

(2b3) Use the histogram to count the gradient and direction of the pixels in the neighborhood window of each key point sequentially, that is, the histogram takes every 10-degree direction as a column, a total of 36 columns, and the direction indicated by the column is the gradient direction of the pixel point, and the column The length of is the gradient amplitude, and the direction represented by the longest column in the histogram is used as the main direction of each key point to complete the direction determination;

(2c) Block the image area around each key point that has completed positioning and direction determination, and calculate the gradient histogram of 8 directions in the 4×4 block centered on the key point, and draw the gradient histogram of each gradient direction Accumulate to generate a unique 128-dimensional vector, use this vector to describe this key point, and obtain the SIFT features of the two images.

Step 3, perform feature matching search to obtain matching points of every two images.

Using the k-d tree algorithm, namely the k-d tree algorithm and the optimal node first algorithm, namely the BBF algorithm, to perform feature matching search on the image, and to realize the feature point matching of the two images, the steps are as follows:

(3a) adopting the k-d tree algorithm to set up a k-d tree for the feature points in the image to be stitched according to the feature points obtained in step (2);

(3b) Use the BBF algorithm to perform feature matching search on the image, and realize the feature point matching of the two images:

(3b1) For each feature point in the input image, find the first two nearest neighbor feature points in the image to be spliced with the closest Euclidean distance in the k-d tree;

(3b2) Calculate the Euclidean distance between the specified feature point and the first nearest neighbor of the two nearest neighbor feature points and the Euclidean distance between the specified feature point and the second nearest neighbor, and compare the ratio with the set ratio threshold 0.49 Compare:

If the ratio is less than the ratio threshold, the specified feature point and the first nearest neighbor point are accepted as a pair of matching points to realize the feature point matching of the two images; otherwise, the specified feature point and the first nearest neighbor point are not accepted as a pair match point.

Step 4, filter the matching points of every two images obtained in step 3 and calculate the optimal transformation matrix H.

This step is performed using the RANSAC algorithm, and the steps are as follows:

(4a) Use the matching point pairs obtained in step 3 as a sample set, and randomly select a RANSAC sample from the sample set, that is, 4 matching point pairs;

(4b) Calculate the current transformation matrix L according to these 4 matching point pairs;

(4c) According to the sample set, the current transformation matrix L and the error metric function, obtain a consistent set C satisfying the current transformation matrix L, and record the number a of elements in the consistent set;

(4d) Set an optimal consistent set, the initial number of elements is 0, compare the number of elements a in the current consistent set with the number of elements in the optimal consistent set: if the number of elements a in the current consistent set is greater than the optimal consistent set number of elements, update the optimal consistent set to the current consistent set, otherwise, do not update the optimal consistent set;

(4e) Calculate the current error probability p:

p＝(1-in_frac ^s ) ^o

Among them, in_frac is the percentage of the number of elements in the current optimal consistent set to the total number of samples in the sample set, s is the minimum number of feature point pairs required to calculate the transformation matrix, the value is s=4, and o is the number of iterations;

(4f) Compare the computed current error probability p with the allowable minimum error probability of 0.01:

If p is greater than the minimum error probability allowed, return to step (4a) until the current error probability p is less than the minimum error probability;

If p is smaller than the allowable minimum error probability, the transformation matrix L corresponding to the current optimal consistent set is the optimal transformation matrix H to be obtained, and the size of the optimal transformation matrix is 3×3.

In step 5, the image is transformed and fused according to the optimal transformation matrix obtained in step 4.

Because the traditional weighted average fusion algorithm is prone to "ghosting" phenomenon, especially in dynamic scenes, when the array camera is used to collect images, if there is a moving object, the direct traditional weighted average fusion algorithm is very ineffective, and it is difficult to reflect the moving object , so the present invention adopts an improved optimal suture algorithm combined with a weighted fusion algorithm to significantly improve the splicing and fusion of the "ghost" phenomenon, and the steps are as follows:

(5a) Transform any one of the two input images according to the transformation matrix obtained in step (4), so that the two images are located in the same coordinate system;

(5b) Correct the brightness of the two images transformed to the same coordinate system in (5a), so that the brightness difference between the two images is the smallest, the process is as follows:

(5b1) Convert both the image to be stitched and the input image into a grayscale image, and calculate the pixel sum of the image to be stitched and the pixel sum of the input image respectively, that is, first calculate the sum g of the pixel values of the non-overlapping part of the input image and the sum of the pixels to be stitched The sum v of the pixel values of the non-overlapping part of the image; then calculate the pixel sum q of the central rectangle of the overlapping area, the height of the central rectangle is 1/2 of the height value of the overlapping area, and the width is 1/1.5 of the width of the overlapping area; then get The pixel sum of the input image is g+q, and the pixel sum of the image to be stitched is v+q;

(5b2) Calculate the ratio b of the pixel sum of the image to be stitched and the input image, and compare the ratio b with 1:

If b is less than 1, multiply the pixel value of each point of the input image by b, and execute (5c);

If b is greater than 1, the pixel value of each point of the image to be stitched is multiplied by the reciprocal of b, and (5c) is executed;

(5c) Find an optimal seam line on the registered and brightness-corrected image:

(5c1) Convert both the input image and the image to be stitched into a grayscale image, and subtract the corresponding pixels of the input image and the image to be stitched in the overlapping area in sequence to obtain the difference image of the overlapping area of the two images, and calculate the difference image The intensity value E(x,y) of each pixel in:

E(x,y)＝|E _gray (x,y)|+E _geometry (x,y),

Among them, E _gray represents the difference between the gray value of the pixels in the overlapping area, and E _geometry represents the difference in the structure value of the pixels in the overlapping area:

E _geometry = (▽x ₁ -▽x ₂ )×(▽y ₁ -▽y ₂ )

Among them, ▽x ₁ -▽x ₂ is the gradient difference in the x direction of the corresponding pixels in the overlapping area between the input image and the image to be stitched,

▽y ₁ -▽y ₂ is the gradient difference in the y direction of the corresponding pixels in the overlapping area between the input image and the image to be stitched;

▽x ₁ is the gradient of each point of the input image in the overlapping area in the x direction, which is obtained by performing a convolution sum operation on the kernel S _x in the x direction and each pixel of the input image in the overlapping area image;

▽ x ₂ is the gradient of each point in the overlapping region of the image to be spliced in the x direction, and the gradient is obtained by the convolution sum of the kernel S _x in the x direction and each pixel of the image to be spliced in the overlapping region;

▽y ₁ is the gradient of each point of the input image in the overlapping area in the y direction, and the gradient is obtained by the convolution sum of the kernel S _y in the y direction and each pixel of the input image in the overlapping area image;

▽y ₂ is the gradient of each point in the overlapping region of the image to be spliced in the y direction, and the gradient is obtained by the convolution sum of the kernel S _y in the y direction and each pixel of the image to be spliced in the overlapping region image;

S _x , S _y are both improved Sobel operator templates, respectively:

(5c2) Using dynamic programming theory, each pixel point in the first line of the difference image is used as the starting point of a suture line, and expanded downwards to find the point with the smallest intensity value among the three adjacent points in the next line, so that it can be used as The expansion direction of the suture line, and so on to the last line, find the suture line with the smallest sum of E(x, y) among all the suture lines generated as the best suture line;

(5d) Perform a weighted average fusion of the rectangles containing the best stitching lines to obtain a stitched panoramic image of the two images:

(5d1) After finding the minimum suture line, take a rectangular area containing the suture line and extending 10 pixels left and right, and perform weighted average of the pixels therein to obtain the fusion map of the rectangular area;

(5d2) Obtain the left part of the rectangular area from the input image, and obtain the right part of the rectangular area from the converted image to be stitched to obtain the final fusion image. So far, the stitching of the two input images is completed.

Step 6, stitching of array images.

(6a) Take the previously spliced image and the image to be spliced as the input two images, continue to repeat steps 2 to 5, and proceed in a loop until the image to be spliced is the third image, and finally get 3 images. Stitch panoramas horizontally;

(6b) Repeat (6a) twice to obtain two horizontal mosaic images, each horizontal mosaic image is stitched by 3 images;

(6c) Determine whether the length and width of the two horizontal mosaic images are multiples of 4: if not, modify the length and width of the two horizontal mosaic images to a multiple of 4, if so, keep the two horizontal mosaic images The length and width of the graph remain unchanged;

(6d) Take the two horizontal mosaic images as input images, rotate them counterclockwise by 90°, repeat steps 2 to 5 to obtain the vertical mosaic images of the two images, and then rotate the spliced vertical images 90° clockwise to finally get Horizontal and vertical stitching panorama of 6 images.

It should be noted that the method of the present invention is not limited to the splicing of six array images, and the method has wide applicability. When there are enough cameras or when the cameras are moved to view separately, more images can be spliced. Complete two, three, four image stitching on the existing basis to meet various needs and apply to various occasions. At the same time, since the array camera can realize simultaneous shooting by six cameras, the accuracy of the stitching result is also guaranteed.

The effect of the present invention can be further illustrated by some experiments.

1. Experimental conditions

The experimental system includes an array camera, as shown in Figure 2. This experiment is carried out under the software environment of VS2010.

2. Experimental content

Using the method of the present invention to collect outdoor images, including a moving person in the scene, the image collected by the array camera is shown in Figure 3, wherein Figure 3 includes six images with overlapping regions, using the method of the present invention to analyze Figure 3 The panorama in Figure 4 is obtained after stitching the six images.

It can be seen from Fig. 4 that the present invention has a good stitching effect on images with dynamic objects, and there is no "ghost" phenomenon; in the whole stitched panorama, no obvious stitching seam is observed; compared with a single image, the stitching Panoramas have a larger field of view and more image details for high-quality stitching results.

Claims (9)

1. a kind of method for panoramic imaging based on camera array, including：

(1) gathered while completing multiple image using array camera, obtain i width images, m≤i；

(2) two images are read in and Scale invariant features transform characteristic point, i.e. SIFT feature is extracted respectively；

(3) characteristic matching lookup is carried out to SIFT feature obtained by step (2), obtains the match point of every two images；

(4) every two images match point obtained by step (3) is screened and calculates optimal transform matrix H；

(5) optimal transform matrix according to obtained by step (4) enters line translation to image, and carries out image co-registration：

(5a) optimal transform matrix according to obtained by step (4) enters line translation to any one width in input two images so that two Width image is located in the same coordinate system and two images have overlapping region；

(5b) carries out gamma correction to the two images that the same coordinate system is transformed in (5a), makes the luminance difference of two images most It is small；

(5c) finds an optimal stitching line on the image of registration；

(5d) is weighted average fusion to the rectangle comprising optimal stitching line, obtains the spliced panoramic image of two images；

(6) array image splices：

The image and image to be spliced that (6a) has once been spliced using before it continue repeat step (2) as the two images of input ~(5), circulation is carried out, and untill image to be spliced is m width images, m ＜=i finally give the horizontally-spliced of m width images Panorama sketch；

(6b) repeats (6a) k times, and k ＜=i and k × m≤i obtain k horizontally-spliced figure, every horizontally-spliced figure is by m width images It is spliced；

(6c) using two width before k width landscape images as input picture, 90 ° of rotate counterclockwise, repeat step (2)~(5) are obtained The longitudinal spliced figure of two images, for remaining k-2 width landscape images, with the image and rotate counterclockwise once spliced before it Image to be spliced after 90 ° repeats step (2)~(5), circulation is carried out, until figure to be spliced as the two images of input As untill being kth width image, then longitudinal image clockwise obtained by splicing is rotated by 90 °, finally gives the transverse and longitudinal of i width images Spliced panoramic figure.

2. according to the method described in claim 1, it is characterised in that：Array camera described in step (1), is by i camera Two-dimensional combination arrangement is carried out in transverse and longitudinal direction, by the position and the focal length that adjust each camera so that these cameras all the time can Embark on journey in column and entire combination is rectangle, be met the image of different requirements.

3. according to the method described in claim 1, it is characterised in that：Two images are read in step (2) and yardstick is extracted respectively Invariant features transform characteristics point, is carried out as follows：

(2a) constructs gaussian pyramid and difference of Gaussian pyramid, detects yardstick spatial extrema；

Metric space extreme value in (2a) as key point, is positioned and direction is determined by (2b) to the key point；

(2c) completes image-region progress piecemeal around the key point that positioning and direction are determined for each, and is in the key point The histogram of gradients in 8 directions is calculated in 4 × 4 block at center, the cumulative of each gradient direction is drawn, generates unique 128 dimension vectors, this key point is depicted with the vector to obtain the SIFT feature of two images.

4. according to the method described in claim 1, it is characterised in that：Characteristic matching in step (3) is searched, and is entered as follows OK：

(3a) treats the characteristic point in stitching image using characteristic point of the k-d tree algorithm according to obtained by step (2) and sets up k-d tree；

(3b) carries out characteristic matching lookup using BBF algorithms to image, realizes the Feature Points Matching of two images：

(3b1) finds out in image to be spliced that Euclidean distance is nearest therewith to each characteristic point in input picture in k-d tree The first two arest neighbors characteristic point；

(3b2) by the Euclidean distance of first arest neighbors in specific characteristic point and two arest neighbors characteristic points and specific characteristic point with The Euclidean distance of second arest neighbors is asked, and the ratio and the proportion threshold value 0.49 of setting are compared：

If ratio is less than the proportion threshold value, it is a pair of match points to receive specific characteristic point and first nearest neighbor point, is realized The Feature Points Matching of two images；Otherwise it is a pair of match points not receive specific characteristic point and first nearest neighbor point.

5. according to the method described in claim 1, it is characterised in that：Described in step (4) to every two width figure obtained by step (3) As match point is screened and calculates optimal transform matrix H, carried out using RANSAC algorithms, its step is as follows：

(4a) using gained matching double points are as sample set in step (3), one RANSAC sample of random selection from sample set, i.e., 4 matching double points；

(4b) calculates current transform matrix L according to this 4 matching double points；

(4c) is met current transform matrix L consistent collection according to sample set, current transform matrix L and error metrics function C, and record the number a of consistent concentration element；

(4d) sets an optimal consistent collection, and finite element number is 0, currently will unanimously concentrate element number a and optimal consistent Element number is concentrated to compare：If current consistent concentration element number a is more than optimal consistent concentration element number, will be optimal Consistent collection is updated to current consistent collection, otherwise, does not then update optimal consistent collection；

(4e) calculates current erroneous Probability p：

P=(1-in_frac^s)^o

Wherein, in_frac is the current optimal consistent percentage for concentrating element number to account for total sample number in sample set, and s is to calculate The minimal characteristic point that transformation matrix needs is to number, and value is s=4, and o is iterations；

The minimum error probability 0.01 that current erroneous Probability p is calculated with allowing is compared by (4f)：

If p is more than the minimum error probability allowed, return to step (4a), until current erroneous Probability p is less than minimal error Untill probability；

If p is less than the minimum error probability allowed, current optimal unanimously to collect corresponding transformation matrix L be required optimal Transformation matrix H.

6. according to the method described in claim 1, it is characterised in that：To transforming to the same coordinate system in (5a) in step (5b) Two images carry out gamma correction, carry out as follows：

Image to be spliced and input picture are converted into gray-scale map by (5b1), calculate respectively image to be spliced pixel and with The pixel of input picture and；

(5b2) calculate image pixel to be spliced and with input image pixels and ratio b；

The input picture pixel value of every is multiplied by (5b3) if the ratio b obtained by being calculated in (5b2) is less than 1 with b；If Ratio b is more than 1, then by the pixel value of every and b reciprocal multiplication of image to be spliced.

7. according to the method described in claim 1, it is characterised in that：In step (5c) one is found on the image of registration most preferably Suture, is carried out as follows：

Input picture and image to be spliced are converted into gray-scale map by (5c1), by input picture and image to be spliced in overlay region The respective pixel in domain is subtracted each other successively, obtains each picture in the error image of two images overlapping region, calculating difference image The intensity level E (x, y) of element：

E (x, y)=| E_gray(x,y)|+E_geometry(x, y),

Wherein, E_grayRepresent the difference of the gray value of overlapping region pixel, E_geometryRepresent the structured value of overlapping region pixel Difference：

E_geometry=(▽ x₁-▽x₂)×(▽y₁-▽y₂)

Wherein, ▽ x₁-▽x₂For input picture and image to be spliced overlapping region respective pixel x directions gradient difference,

▽y₁-▽y₂For input picture and image to be spliced overlapping region respective pixel y directions gradient difference；

▽x₁Each put the gradient in x directions in overlapping region for input picture, the gradient by x directions core S_xWith input picture The computing that each pixel in the image of overlapping region does convolution sum is obtained；

▽x₂Each put the gradient in x directions in overlapping region for image to be spliced, the gradient by x directions core S_xWith it is to be spliced The computing that each pixel of the image in the image of overlapping region does convolution sum is obtained；

▽y₁Each put the gradient in y directions in overlapping region for input picture, the gradient by y directions core S_yWith input picture The computing that each pixel in the image of overlapping region does convolution sum is obtained；

▽y₂Each put the gradient in y directions in overlapping region for image to be spliced, the gradient by y directions core S_yWith it is to be spliced The computing that each pixel of the image in the image of overlapping region does convolution sum is obtained；

The S_x, S_yIt is improved Sobel operators template, is respectively：

<mrow> <msub> <mi>S</mi> <mi>x</mi> </msub> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <mo>-</mo> <mn>3</mn> </mrow> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>3</mn> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>-</mo> <mn>3</mn> </mrow> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>3</mn> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> <msub> <mi>S</mi> <mi>y</mi> </msub> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <mo>-</mo> <mn>3</mn> </mrow> </mtd> <mtd> <mrow> <mo>-</mo> <mn>1</mn> </mrow> </mtd> <mtd> <mrow> <mo>-</mo> <mn>3</mn> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>0</mn> </mtd> </mtr> <mtr> <mtd> <mn>3</mn> </mtd> <mtd> <mn>1</mn> </mtd> <mtd> <mn>3</mn> </mtd> </mtr> </mtable> </mfenced> <mo>;</mo> </mrow>

(5c2) uses the thoery of dynamic programming, using each pixel of error image the first row as a suture starting point, to Lower extension, finds the minimum point of intensity level in three adjacent points of next line, is allowed to the propagation direction as suture, successively class Shift last column onto, a minimum suture of E (x, y) sum is found out in all sutures of generation as optimal suture Line.

8. method according to claim 3, it is characterised in that：Construction gaussian pyramid and Gaussian difference parting in step (2a) Word tower, detects yardstick spatial extrema, carries out as follows：

(2a1) calculates pyramidal number of plies n according to the original size of image and the size of tower top image：

N=log₂{min(M,N)}-t,t∈[0,log₂{min(M,N)})

Wherein M, N are respectively the length and width of original image, and t is the logarithm value of the minimum dimension of pyramid tower top image；

(2a2), using original image as the first layer of Gauss pyramid, then to original image, successively depression of order is sampled, each down-sampled institute Obtained new images are pyramidal new one layer, untill n-th layer, obtain a series of descending images, from top to bottom Tower-like model is constituted, initial pictures pyramid is obtained；

One image of every layer of initial pictures pyramid is done Gaussian Blur by (2a3) using different parameters so that pyramidal every Layer contains multiple Gaussian Blur images, and every layer of multiple image of pyramid is collectively referred to as into one group, gaussian pyramid is obtained；

Adjacent two image subtractions up and down, obtain Gaussian difference parting word in every group of the gaussian pyramid that (2a4) obtains (2a3) Tower；

(2a5) takes each pixel in every group of difference of Gaussian pyramid, respectively two with this and up and down by them All pixels point is made comparisons in 26 neighborhoods, if the value of the pixel taken from difference of Gaussian pyramid is maximum or minimum Value, then the pixel value of taken point is a metric space extreme value of the image under current scale, and wherein metric space is by Gauss gold The realization of word tower, each group of the corresponding different yardstick of each image.

9. method according to claim 3, it is characterised in that：Crucial point location is carried out in step (2b) and direction is determined, Carry out as follows：

(2b1) removes the point of low contrast by interpolation, and eliminates skirt response, and completion is accurately positioned to key point；

(2b2) gathers pixel in its σ neighborhood window of place gaussian pyramid image 3 for pinpoint key point in (2b1) Gradient and directional spreding feature, the modulus value of gradient and direction are as follows：

<mrow> <mi>m</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>)</mo> </mrow> <mo>=</mo> <msqrt> <mrow> <msup> <mrow> <mo>(</mo> <mi>L</mi> <mo>(</mo> <mrow> <mi>x</mi> <mo>+</mo> <mn>1</mn> <mo>,</mo> <mi>y</mi> </mrow> <mo>)</mo> <mo>-</mo> <mi>L</mi> <mo>(</mo> <mrow> <mi>x</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>y</mi> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>+</mo> <msup> <mrow> <mo>(</mo> <mi>L</mi> <mo>(</mo> <mrow> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>+</mo> <mn>1</mn> </mrow> <mo>)</mo> <mo>-</mo> <mi>L</mi> <mo>(</mo> <mrow> <mi>x</mi> <mo>,</mo> <mi>y</mi> <mo>-</mo> <mn>1</mn> </mrow> <mo>)</mo> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> </msqrt> </mrow>

θ (x, y)=tan^-1((L(x,y+1)-L(x,y-1))/L(x+1,y)-L(x-1,y)))

Wherein, L is the metric space value where key point, and the modulus value m (x, y) of gradient is added by σ=1.5 σ _ oct Gaussian Profile Into the 3 σ principles sampled by yardstick, neighborhood windows radius is 3 × 1.5 σ _ oct；

(2b3) counts the gradient of pixel and direction in each crucial vertex neighborhood window using histogram successively, i.e., by histogram with Every 10 degree of directions are a post, totally 36 posts, and the direction that post is represented is pixel gradient direction, and the length of post is gradient magnitude, The direction that most long column is represented using in histogram completes direction and determined as the principal direction of each key point.