CN102855649B - Method for splicing high-definition image panorama of high-pressure rod tower on basis of ORB (Object Request Broker) feature point - Google Patents

Abstract

The invention discloses a panorama splicing method for high-definition images of high-voltage pole towers based on ORB feature points: 1: Read ultra-high-resolution high-voltage pole-tower images, the size of which is W×H, and use bilinear interpolation to super-high-resolution The image is sampled and reduced to obtain a w×h image, where W, H, w, and h are integers greater than 0, k is an integer greater than 0; 2: use the ORB algorithm to extract features from all sampled images; 3: perform rough matching on the ORB features extracted in the second step; 4: use the matching point pairs extracted in the previous step The ORB feature is extracted again from the image block where the matching point pair of the high-resolution image is located, and the precise matching is performed; five: through the matching point pair obtained above, calculate the transformation matrix H ₀ between adjacent images; six: use gradual and gradual A method for fusion of super-high resolution adjacent images. It realizes the seamless splicing of ultra-high resolution images, reduces the time required for splicing, improves the splicing efficiency, and has a good beneficial effect especially on high-definition images.

Description

Panoramic mosaic method of high-voltage tower high-definition images based on ORB feature points

technical field

The invention relates to a high-voltage tower high-definition panoramic splicing method, in particular to a high-voltage tower high-definition panoramic splicing method based on ORB feature points.

Background technique

In recent years, the sustained and rapid development of my country's national economy has put forward higher and higher requirements for my country's electric power industry. Due to the vast territory of our country and the complex terrain of transmission line corridors, the limitations of the traditional manual line inspection operation mode have become increasingly prominent. It has become possible to use unmanned aerial vehicles equipped with digital imaging equipment to conduct detailed inspections of overhead transmission lines. Although the high-definition rate has reached the requirement of seeing the fittings of the high-voltage transmission line in detail, due to the small field of view of the imaging equipment, the collected high-definition images cannot include all the equipment of the high-voltage tower.

Panoramic image stitching technology has a wide range of application values in satellite remote sensing detection, meteorology, medicine, military, aerospace, large-area cultural heritage protection, and virtual scene realization. The high-voltage poles and towers of overhead transmission lines have large-scale image characteristics, and it is impossible to capture panoramic and ultra-high-resolution images at one time with ordinary digital imaging equipment. The above problems can be successfully solved by using image stitching technology, and the synthesis of ultra-high-resolution high-voltage tower images has been successfully realized.

Panoramic image stitching technology can stitch multiple images collected by digital imaging equipment into a panoramic image with a larger field of view, and the final panoramic image has less distortion, and the regions of interest are all displayed on a panoramic image. Panoramic image stitching technology mainly involves three aspects: feature point extraction, feature point matching and image fusion technology, in which the effect of feature point extraction directly affects the effect of later image stitching.

At present, SIFT and SURF are relatively popular feature point extraction methods, although the above two feature point extraction methods have been relatively mature in image stitching and many other aspects. However, when extracting feature points from high-resolution images, a lot of time will be spent on feature point extraction.

Contents of the invention

In order to solve the above-mentioned problems in image stitching, the present invention proposes a panoramic stitching method for high-voltage tower high-definition images based on ORB feature points with low time complexity and good stitching effect. It uses the feature point matching algorithm combining rough matching and precise matching to realize the seamless stitching of ultra-high resolution images, reduce the time required for stitching, and improve the efficiency of stitching, especially for high-definition images. .

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for panoramic stitching of high-voltage tower high-definition images based on ORB feature points, the specific steps are:

The first step: read the ultra-high-resolution high-voltage tower image, and sample and reduce the image;

The second step: use the ORB algorithm to extract features from all the images after sampling reduction;

The third step: use the extracted ORB features to perform nearest neighbor matching, and filter the obtained matching point pairs through the RASANC algorithm to obtain rough matching point pairs;

Step 4: Use the coordinates of the coarse matching point pair extracted in the previous step to calculate the corresponding coordinates in the original ultra-high resolution image, and extract the ORB feature again in the image block where the matching point pair of the original ultra-high resolution image is located , for an exact match;

Step 5: Calculate the transformation matrix H ₀ between adjacent images;

Step 6: Use the gradual in and gradual out method to fuse the super high resolution adjacent images to obtain the super high resolution panoramic image, and the splicing ends.

In the first step, the sampling reduction method is as follows: use bilinear interpolation method to sample and reduce the ultra-high resolution image to be spliced, the original image size is W×H, and the obtained image size is w×h, where W , H, w, h are integers greater than 0, k is an integer greater than 0.

The specific steps of the feature extraction in the second step are:

(2-1) Perform Oriented FAST feature point detection:

(2-2) Generate Rotated BRIEF feature descriptor;

Randomly select several point pairs near the feature points, combine the gray values of these point pairs into a binary string, and use this binary string as the feature descriptor of the feature point.

The concrete process of described step (2-2) is:

a Generate a BRIEF feature descriptor;

b generate the Rotated BRIEF feature descriptor;

The direction vector extracted from the Oriented FAST algorithm is added to the BRIEF feature, rotated, and the oriented Brief is obtained, which is called Steered Brief; the greedy learning algorithm is used to screen out the steered brief with high variance and high irrelevance, and the result is called Steered Brief. For rBRIEF; calculate the distance between each SBRIEF and 0.5, and create a container T; put the first SBRIEF into the result container R, and remove it from container T; take the next SBRIEF from container T, and and the result container All SBRIEFs in R are compared, if their correlation is less than a certain threshold, they are added to the result container R, otherwise they are discarded;

Repeat step b until there are 256 SBRIEFs in the result container R, if there are less than 256 SBRIEFs in the result container R, change the threshold and repeat the above steps.

The specific steps of the third step are as follows:

(3-1) Select LSH as the nearest neighbor matching point pair calculation;

(3-2) Use the RASANC algorithm to screen the coarse matching point pairs generated in step (3-1), select the matching point pairs that meet the requirements, that is, the inliers, and delete the wrong matching point pairs.

The concrete process of described step (3-2) is:

(a) Interior point initialization: Randomly select 4 pairs of matching points in a given matching point pair;

(b) Calculate the transformation matrix H ₀ through interior points;

(c) For the remaining matching point pairs in the matching point pair, calculate the distance between them and the transformation matrix H ₀ , if the result is less than a certain threshold, add it to the interior point set, and according to the new interior point set, use The least square method updates the transformation matrix H ₀ , otherwise continue to judge the remaining matching point pairs;

(d) Repeat step (c) until the number of interior points no longer increases.

The concrete steps of described 4th step are as follows:

(4-1) Let M _l and M _r be the super-high resolution original two adjacent images, m _l and m _r are the two adjacent images after sampling and reduction respectively, and the coordinates of the coarse matching point pair calculated in the third step Respectively (x _li , y _li ) and (x _rj , y _rj ) where 0≤i, j≤n, n is the number of pairs of matching points sought;

(4-2) Taking (X _li , Y _yli ) and (X _rj , Y _rj ) as the center respectively, the range image blocks with γ as the radius are respectively I _l and I _r , where:

(4-3) Extracting ORB features from image blocks I _l and I _r respectively;

(4-4) find the matching point pair of I _l and I _r ;

(4-5) For all matching point pairs, repeat the above steps to generate exact matching matching point pairs.

The concrete steps of described sixth step are as follows:

(6-1) Transform the corresponding image according to the transformation matrix H ₀ between the images, and determine the overlapping area between the images;

(6-2) Let I _l and I _r be two adjacent images respectively, and I is the fused image:

I(x,y)=(1-τ(k))×I _l (x,y)+τ(k)×I _r (x,y)+d (1)

Where 0≤d≤1 is the fine-tuning coefficient, 0≤τ(k)≤1 is the weighting function,

$\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, m is the width of the overlapping area, and k is the number of pixels away from the leftmost of the overlapping area. In this way, the larger the overlapping area, the smoother τ(k) will be, so that the transition between images can be smooth.

Beneficial effects of the present invention:

1. The feature points extracted by the ORB algorithm have a good effect in image stitching applications, and its operation time is two orders of magnitude faster than the SIFT algorithm and one order of magnitude faster than the SURF algorithm;

2. By extracting ORB feature points in the reduced image, the problem of time-consuming and insufficient memory caused by too many feature points in the ultra-high resolution image is solved;

3. In the ultra-high resolution image, the ORB algorithm is used to extract the feature points in the image block where the matching point is located, and perform accurate matching, which solves the error in the image stitching after sampling;

4. Using the method of combining rough matching and precise matching, the matching time has been significantly reduced, and the matching accuracy has also been greatly improved;

5. From the perspective of the entire ultra-high resolution image stitching, the stitching speed has been greatly improved.

Description of drawings

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

Figures 2 and 3 are images before splicing;

Figures 4 and 5 show the effect after splicing.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, the method steps of the present invention are as follows:

The first step is to read the ultra-high-resolution high-voltage tower image with a size of W×H, and use the bilinear interpolation method to sample and reduce the ultra-high-resolution image to be stitched to obtain a w×h image, where W, H, w , h is an integer greater than 0, k is an integer greater than 0;

The second step is to use the ORB algorithm to perform feature extraction on all sampled images:

The ORB feature uses the Oriented FAST feature point detection operator and the Rotated BRIEF feature descriptor. The ORB algorithm not only has the detection effect of SIFT features, but also has the characteristics of rotation, scale scaling, and brightness change invariance. The most important thing is that its time complexity is greatly reduced compared with SIFT, making ORB algorithm in high-definition image stitching And real-time video image stitching has a great application prospect.

Specifically include the following steps:

2-1) Oriented FAST feature point detection:

The present invention adds direction vectors on the basis of FAST feature point detection to make it have directionality.

a) Use the FAST feature point detection algorithm to quickly detect key points;

b) Find the centroid and direction of the block where the key point is located:

The block centroids are extracted as follows:

m _pq =∑ _{x, y} x ^p y ^q I(x, y) (1)

$\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\frac{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 10</span>}}}{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 00</span>}}}<span\; class="notranslate">\; ,</span>\frac{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 01</span>}}}{{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 00</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, m _pq is the moment of the block, p, q ∈ (0, 1), x, y ∈ block, C is the centroid of the block to be sought; the extraction of the block direction:

θ=atan2(m ₀₁ , m ₁₀ ) (3)

In this way, the directional FAST key points have been extracted. However, FAST feature point detection cannot handle multi-scale images, but the original image can be made into a pyramid, and then the above steps are performed on each image, so that Oriented FAST supports multi-scale changes.

2-2) Generate Rotated BRIEF feature descriptor

The main idea of BRIEF is to randomly select several point pairs near the feature points, combine the gray value of these point pairs into a binary string, and use this binary string as the feature descriptor of the feature point. Its advantage is that the calculation speed is very fast.

a) Generate the BRIEF feature descriptor:

Given a picture;

Smooth the image to reduce image noise;

Select an area block p of SXS pixels on the image, where 5≤S≤15, and extract the BRIEF feature on p:

Define the τ test, and the calculation formula of the τ test is as follows:

$\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; 1</span>}& \mathrm{<span\; class="notranslate">\; if\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

x, y are two-pixel positions within p, that is, x and y are two-dimensional coordinates of the shape [u, v], and p(x) and p(y) are the brightness values of x and y pixels;

A BRIEF feature is a binary string composed of several τ tests, construct a specific [x,y] pair, and obtain the BRIEF feature according to the following formula

${\mathrm{<span\; class="notranslate">\; f</span>}}_{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; :</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ;</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

b) Generate Rotated BRIEF feature descriptor:

The direction vector extracted from the Oriented FAST algorithm is added to the BRIEF feature, and rotated to obtain an oriented BRIEF, which is called Steered BRIEF;

Use a greedy learning algorithm to screen steered briefs with high variance and high irrelevance. The result is called rBRIEF. The algorithm is as follows:

Calculate the distance between each SBRIEF and 0.5, and create a container T;

Among them, the greedy algorithm:

Put the first SBRIEF into the result container R and remove it from the container T;

Take out the next SBRIEF from container T, and compare it with all SBRIEFs in the result container R, if its correlation is less than a certain threshold, add it to the result container R, otherwise discard;

Repeat step b) until there are 256 SBRIEFs in the result container R, if there are less than 256 SBRIEFs in the result container R, then change (increase) the threshold, and repeat the above steps;

The third step is to roughly match the ORB features extracted in the second step. The specific steps are as follows:

3-1) The rBRIEF feature is a binary form feature, and Locality Sensitive Hashing (LSH) is selected as the nearest matching point pair for calculation;

3-2) Use the RASANC algorithm to screen the matching point pairs generated in the previous step:

RANSAC is the abbreviation of "RANdom SAmple Consensus (Random Sampling Consensus)". It is a robust estimation method proposed by Fischler and Bolles in 1981, which can estimate high-precision parameters from a large number of outlier data sets. The basic idea is to design a search engine when estimating parameters, use this search engine to iteratively filter out input data consistent with the estimated parameters, and then use these data to estimate parameters.

The present invention utilizes RANSAC algorithm to screen all matching point pairs, selects matching point pairs that meet the requirements of the parameter model, that is, interior points, and deletes wrong matching point pairs. The specific algorithm is as follows:

Interior point initialization: Randomly select 4 pairs of matching points in a given matching point pair;

Calculate the transformation matrix H ₀ through the interior point;

For the remaining matching point pairs in the matching point pair, calculate the distance between them and the transformation matrix H ₀ , if the result is less than a certain threshold, add it to the interior point set, and use the least squares method according to the new interior point set Update the transformation matrix H ₀ , otherwise continue to judge the remaining matching point pairs;

Repeat the previous step until the number of interior points no longer increases.

The fourth step, through the above three steps, the transformation matrix H ₀ between adjacent images after sampling can be calculated, and images can be spliced by a certain fusion method, but the current transformation matrix H ₀ is calculated in the sampled image If it is used directly in the ultra-high resolution original image, the overlapping area of the image after splicing may not be very accurate, and there will be errors in pixel coordinates at the sampling level. The present invention uses the matching point pairs extracted in the previous step to extract the ORB feature again from the image block where the matching point pair of the original ultra-high resolution image is located, and then performs precise matching, so as to solve the above-mentioned problems. Specific steps are as follows:

4-1) Let M _l and M _r be the original super-resolution two adjacent images, m _l and m _r are the two adjacent images after sampling and reduction respectively, and the coordinates of matching point pairs calculated in the third step are respectively (x _li , y _li ) and (x _rj , y _rj ), where 0≤i, j≤n, n is the number of pairs of matching points sought;

4-2) Taking (X _li , Yy _li ) and (X _rj , Y _rj ) as centers respectively, the range image blocks with γ as radius are I _l and I _r respectively, where γ is any range of 9≤γ≤100 Integers of , where:

4-3) Extract ORB features from image blocks I _l and I _r respectively;

4-4) find the matching point pair of I _l and I _r ;

4-5) For all matching point pairs, repeat the above steps to generate exact matching matching point pairs.

The fifth step is to calculate the transformation matrix H ₀ between adjacent images through the matching point pairs obtained above;

The sixth step is to use the gradual in and gradual out method to fuse the super high resolution adjacent images, the specific steps are as follows:

6-1) According to the transformation matrix H ₀ between the images, the corresponding images can be transformed to determine the overlapping area between the images;

6-2) Let I _l and I _r be two adjacent images respectively, and I is the fused image:

I(x,y)=(1-τ(k))×I _l (x,y)+τ(k)×I _r (x,y)+d (1)

Where 0≤d≤1 is the fine-tuning coefficient, 0≤τ(k)≤1 is the weighting function,

$\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, m is the width of the overlapping area, and k is the number of pixels away from the leftmost of the overlapping area. In this way, the larger the overlapping area, the smoother τ(k) will be, so that the transition between images can be smooth.

In order to test the beneficial effect of the present invention in ultra-high resolution, in the experiment, the present invention tests the splicing effect of 5 groups of ultra-high-resolution high-voltage tower images. The experimental conditions are as follows: CPU is Intel Core i32.27GHz, 2G memory. Table 1 compares ORB with the popular feature point detection methods in terms of time complexity. The image size used is 5184X3456. It can be seen from Table 1 that the time complexity of ORB feature point detection algorithm is significantly better than SURF and SIFT The feature detection algorithm is nearly an order of magnitude faster than SURF, and nearly two orders of magnitude faster than SIFT; Table 2 shows the time complexity of using the ORB algorithm directly on the original ultra-high resolution image and the first rough and then fine block stitching algorithm. From the above comparison, it can be seen from the table that the stitching method of the present invention is nearly 150% faster than the method of stitching directly using the ORB algorithm on the original ultra-high resolution image. Combining the above two experimental results, the present invention not only improves the detection efficiency in the feature point inspection stage, but also greatly reduces the time complexity, and in terms of overall splicing, the present invention uses a block image splicing method that is coarse first and then refined, which also greatly reduces the splicing time. The time complexity greatly improves the stitching efficiency for the stitching of ultra-high resolution images.

Table 1. Comparison of detection time of each feature point detection algorithm

Table 2. Comparison of direct splicing and block splicing time

Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, those skilled in the art do not need to pay creative work Various modifications or variations that can be made are still within the protection scope of the present invention.

Claims (5)

1., based on the clear image panorama joining method of high pressure stem tower height of ORB unique point, it is characterized in that, concrete steps are:

The first step: read ultrahigh resolution high pressure shaft tower image, and sampling is carried out to image reduce;

Second step: all imagery exploitation ORB algorithms after reducing sampling carry out feature extraction;

The concrete steps of the feature extraction in described second step are:

(2-1) Oriented FAST feature point detection is carried out:

(2-2) Rotated BRIEF Feature Descriptor is generated;

Near unique point, the some points of random selecting are right, and these size groups putting right gray-scale value are synthesized a binary string, and using the Feature Descriptor of this binary string as this unique point;

3rd step: utilize the ORB feature extracted to carry out most proximity matching, by RASANC algorithm, the matching double points obtained is screened, obtain thick matching double points;

The concrete steps of described 3rd step are as follows:

(3-1) LSH is selected to calculate most proximity matching point pair;

(3-2) utilize RASANC algorithm to be screened by the matching double points that step (3-1) generates, select and meet the requirements of matching double points i.e. interior point, deletion error matching double points;

4th step: the thick matching double points coordinate that in utilization, step is extracted, calculates the respective coordinates in original ultrahigh resolution image, and again extract ORB feature in the image block at the matching double points place of original ultrahigh resolution image, carry out exact matching;

The concrete steps of described 4th step are as follows:

(4-1) M is made _land M _rfor original two adjacent images of ultrahigh resolution, m _land m _rbe respectively two adjacent images after reducing of sampling, the thick matching double points coordinate that the 3rd step calculates is respectively (x _li, y _li) and (x _rj, y _rj) wherein 0≤i, j≤n, n be required match point logarithm;

(4-2) respectively with (X _li, Y _li) and (X _rj, Y _rj) centered by, be that the range image block of radius is respectively I with γ _land I _r, wherein:

(4-3) at image block I _l, I _rthe middle ORB of extraction respectively feature;

(4-4) I is obtained _land I _rmatching double points;

(4-5) to all matching double points, repeat above step, generate the matching double points of exact matching;

5th step: calculate the transformation matrix H between adjacent image ₀;

6th step: utilize to be fade-in and gradually go out method and merge ultrahigh resolution adjacent image, obtain ultrahigh resolution panoramic picture, splicing terminates.

2. the clear image panorama joining method of a kind of high pressure stem tower height based on ORB unique point as claimed in claim 1, it is characterized in that, in the described first step, the method reduced of sampling is: utilize bilinear interpolation that ultrahigh resolution image to be spliced is carried out sampling and reduce, original image is of a size of W × H, the picture size obtained is w × h, wherein W, H, w, h be greater than 0 integer q is for being greater than 0 integer.

3. the clear image panorama joining method of a kind of high pressure stem tower height based on ORB unique point as claimed in claim 1, it is characterized in that, the detailed process of described step (2-2) is:

A generates BRIEF Feature Descriptor;

B generates Rotated BRIEF Feature Descriptor;

The direction vector extracted in Oriented FAST algorithm is joined in BRIEF feature, rotates, obtain oriented BRIEF, be referred to as Steered BRIEF; Have high variance and the incoherent Steered BRIEF of height with greedy learning algorithm screening, result is referred to as rBRIEF; Calculate the distance of each SBRIEF and 0.5, and create container T; First SBRIEF is put into result container R, and removes from container T; From container T, take out next SBRIEF, and compare with all SBRIEF in result container R, if its correlativity is less than certain threshold value, then adds in result container R, otherwise abandon;

Repeat step b until have 256 SBRIEF in result container R, if be less than 256 SBRIEF in result container R, then change threshold value, and repeat above step.

4. the clear image panorama joining method of a kind of high pressure stem tower height based on ORB unique point as claimed in claim 1, it is characterized in that, the detailed process of described step (3-2) is:

Point initialization in (a): randomly draw 4 pairs of matching double points in given matching double points;

B () calculates transformation matrix H by interior point ₀;

C (), to matching double points remaining in matching double points, calculates they and transformation matrix H ₀distance, if result is less than certain threshold value, then joined in interior set, and according to some set in new, use least square method to upgrade transformation matrix H ₀, otherwise continue to judge remaining matching double points;

D () repeated execution of steps (c), until interior some number no longer increases.

5. the clear image panorama joining method of a kind of high pressure stem tower height based on ORB unique point as claimed in claim 1, it is characterized in that, the concrete steps of described 6th step are as follows:

(6-1) according to the transformation matrix H between image ₀, corresponding image is converted, determines the overlapping region between image;

(6-2) I is made _land I _rbe respectively two adjacent images, I is the image after merging:

I(x，y)＝(1-τ(k))×I _l(x，y)+τ(k)×I _r(x，y)+d (1)

Wherein 0≤d≤1 is fine setting coefficient, and 0≤τ (k)≤1 is weighting function,

$\mathrm{\&tau;}\left(k\right)=\frac{k}{m}---\left(2\right)$

Wherein m is overlapping peak width, and k is leftmost pixel count from overlapping region, and such overlapping region is larger, and τ (k) will be milder, makes seamlessly transit between image.