CN106683046B - Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining - Google Patents

Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining Download PDF

Info

Publication number
CN106683046B
CN106683046B CN201610954653.2A CN201610954653A CN106683046B CN 106683046 B CN106683046 B CN 106683046B CN 201610954653 A CN201610954653 A CN 201610954653A CN 106683046 B CN106683046 B CN 106683046B
Authority
CN
China
Prior art keywords
image
points
feature
point
distance
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201610954653.2A
Other languages
Chinese (zh)
Other versions
CN106683046A (en
Inventor
王帅
张莹莹
刘向阳
张江州
姜树明
阎淮海
张元元
魏志强
王文爽
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong public safety inspection and Testing Technology Co.,Ltd.
Original Assignee
INFORMATION RESEARCH INSTITUTE OF SHANDONG ACADEMY OF SCIENCES
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by INFORMATION RESEARCH INSTITUTE OF SHANDONG ACADEMY OF SCIENCES filed Critical INFORMATION RESEARCH INSTITUTE OF SHANDONG ACADEMY OF SCIENCES
Priority to CN201610954653.2A priority Critical patent/CN106683046B/en
Publication of CN106683046A publication Critical patent/CN106683046A/en
Application granted granted Critical
Publication of CN106683046B publication Critical patent/CN106683046B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4038—Image mosaicing, e.g. composing plane images from plane sub-images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/90—Dynamic range modification of images or parts thereof
    • G06T5/92—Dynamic range modification of images or parts thereof based on global image properties
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20024—Filtering details
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20212—Image combination
    • G06T2207/20221—Image fusion; Image merging

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a real-time image splicing method for reconnaissance and evidence collection of a police unmanned aerial vehicle, which comprises three steps of improved ORB algorithm, image registration and image fusion; the method comprises the steps of firstly constructing a multi-scale space, obtaining an optimal corner detection threshold value by using a significance analysis model, extracting feature points, describing the feature points by using an ORB descriptor, and finally realizing rapid matching by using a Hamming distance combined RANSAC method.

Description

Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining
Technical Field
The invention belongs to the technical field of police unmanned aerial vehicle shooting and evidence obtaining, and particularly relates to a real-time image splicing method for reconnaissance and evidence obtaining of a police unmanned aerial vehicle.
Background
In recent years, the increasing rate of computer crime at home and abroad poses serious threats to national security and social security, serious losses to legal public and private property, and new challenges and requirements on computer evidence-taking technology. Computer evidence collection is used as a new research field, and has great significance for crime fighting and social stability maintenance. The rotary wing type unmanned aerial vehicle is applied to reconnaissance and evidence obtaining, has the advantages of fast response, high real-time performance, real and reliable images and the like, and can effectively solve the problems of insufficient reconnaissance and evidence obtaining means and low efficiency. However, as the aerial image of the unmanned aerial vehicle has a large amount of information and a plurality of visual angles, certain challenges are brought to subsequent information analysis.
The unmanned aerial vehicle is used for reconnaissance and evidence obtaining, and in order to timely and accurately reflect the field situation, the obtained images need to be spliced in real time. The accuracy and efficiency of image feature point matching affect the quality of image stitching. At present, many algorithms are applied to matching of image feature points. The SIFT algorithm is used as a classic feature point matching algorithm, although the matching accuracy is high, the calculation complexity is high, and the real-time requirement cannot be met. Then Bay et al improve it and propose an algorithm for SURF feature point extraction. In recent years, many new feature point matching algorithms, such as BRIEF, ORB, BRISK, FREAK, etc., have emerged.
ORB is an algorithm based on FAST feature extraction and BRIEF feature description, has the advantage of high speed, but does not have scale invariance, and in the feature extraction stage, the algorithm threshold is selected fixedly, and the difference of significant features between images is not considered. Therefore, the method has extremely strong application value to the improvement of the ORB algorithm.
Disclosure of Invention
The invention aims at the problems and provides a real-time image splicing method for police unmanned aerial vehicle reconnaissance and evidence collection.
In order to achieve the technical purpose, the invention adopts a real-time image splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining, which comprises three steps of improved ORB algorithm, image registration and image splicing;
the improved ORB algorithm includes: the ORB is improved, a significant model based on space-frequency domain analysis is applied to the selection of the optimal threshold value in the characteristic extraction stage by combining with a KSW entropy method, and a multi-scale space is constructed by utilizing a Gaussian pyramid;
inputting an image I, firstly converting the image I from an RGB color space to a CIE L ab color space, wherein the image comprises three channels of a L channel, an a channel and a b channel in a L ab color space, wherein a L channel is a brightness channel, and the a channel and the b channel are color channels, and for the color channels, eliminating fine texture details by adopting Gaussian blur in a time domain to obtain a feature map of the corresponding channel;
any input image can be represented by a magnitude spectrum and a phase spectrum, wherein the phase spectrum contains texture detail information of the image, the magnitude spectrum contains light and shade contrast information, if only the phase spectrum is reserved, the obtained image significant features contain partial background information and interfere the detection of image feature points to cause a mismatching phenomenon, a high-pass filter can realize the purpose of sharpening the edge of a target and maximally reserve edge information by attenuating and inhibiting low-frequency components, and in order to enhance the image detail information, sharpen the edge of the target and reduce noise interference, a high-frequency enhanced Butterworth high-pass filter is adopted;
cut-off frequency of D0The nth order butterworth high pass filter is defined as:
Figure GDA0002264094360000031
wherein the content of the first and second substances,
Figure GDA0002264094360000032
representing the frequency midpoint (u, v) and the frequency rectangle center
Figure GDA0002264094360000033
The distance of (d);
inputting an image I, and combining feature maps of L, a and b to obtain a final significant feature SM as follows:
SM(x,y)＝||IL-IL(x,y)||+||Ia-Ia(x,y)||+||Ib-Ib(x,y)||
wherein, IL(x, y) is the corresponding characteristic image pixel value of the brightness channel of the original image after passing through a high-frequency enhanced Butterworth high-pass filter, Ia(x,y),Ib(x, y) are respectively the corresponding characteristic map pixel values of the color channels, IL,Ia,IbRespectively, the average feature vectors of the corresponding channel images, | | | · | |, which is a two-norm Euclidean distance;
the selection of the threshold value should reasonably change along with the change of the image gray scale characteristics, and the selection of the optimal threshold value is determined by combining a KSW entropy method according to the image significant characteristics, and the specific steps are as follows:
let the image gray scale range be [0, L-1]The threshold t divides the data into A, B categories, and the corresponding probability distributions are { p } respectively0,p1,p2,...,pt},{pt+1,pt+2,p2,...,pL-1In which p isiFor the frequency of occurrence of the corresponding gray level, order
Figure GDA0002264094360000034
Then the entropy for A, B classes would be:
Figure GDA0002264094360000035
Figure GDA0002264094360000036
the total entropy of the image is H ═ HA+HB. The optimal threshold value T is
Figure GDA0002264094360000037
K is a proportionality coefficient, and since the threshold of the feature point is related to the pixel contrast of the image, the gray level difference of the image, namely the optimal threshold, is determined by the significant feature of the image and an entropy method;
extracting feature points by using a best threshold obtained by combining a Fast corner detection algorithm, and describing the feature points by using an rBRIEF descriptor for subsequent image registration;
the image registration comprises the steps of utilizing feature point matching and utilizing RANSAC algorithm to screen matching points;
the characteristic point matching is used for searching two characteristic points with the shortest distance in two groups of characteristic point sets by using a distance function, and the distance between two binary descriptors can be represented by using a Hamming distance, wherein the Hamming distance refers to the number of characters with different corresponding positions between two character strings with the same length. The smaller the Hamming distance is, the more similar the two binary descriptors are;
calculating the shortest Hamming distance and the next shortest Hamming distance of each feature point to obtain a group of feature point matching pairs, and considering that the two feature points are matched when the ratio of the shortest distance to the next shortest distance is less than a threshold value;
the method comprises the steps of selecting a certain number of samples at random to estimate model parameters by screening matching points by using a RANSAC algorithm, classifying the rest data according to the estimated parameters, determining that one part of data is within an allowable error range, namely an inner point, and removing an error matching point pair through multiple hypothesis verification if the part of data is an outer point;
the RANSAC algorithm needs to use a homography matrix H which describes the transformation relation between the coordinates of two image points, including translation, rotation, scaling and the like, and can find the position of a point in one image in the other image through the matrix H, and assume a pair of matching points p in an image 1 and an image 21(x,y)，p2The transformation relationship between (x ', y') is:
Figure GDA0002264094360000041
8 parameters of the matrix H can be calculated by 4 pairs of matching points, and the RANSAC algorithm comprises the following steps:
(1) setting an initial value of iteration times as 0, a maximum iteration time N, an internal point number threshold T1 and an error threshold T2;
(2) randomly selecting 4 pairs from a plurality of pairs of points to be matched, and calculating parameters of a transformation matrix H between two images;
(3) calculating the distance between the coordinate values of the other feature points after H transformation and the matching points of the feature points, if the distance is smaller than an error threshold value T2, considering the matching point pair as an inner point, and if the distance is not smaller than the error threshold value T2, calculating the number of the inner points;
(4) if the number of inliers is greater than the threshold number of inliers T1, the current model is saved as the optimal model. Otherwise, adding one to the iteration times, and turning to the step (2) to continue the next iteration;
(5) if the maximum iteration number N is reached, returning a group of interior points with the maximum number of corresponding interior points, and obtaining a transformation matrix H;
the image stitching means: after the images are registered, the images are spliced through image fusion, wherein the image fusion is the last step of image splicing and mainly comprises two parts: merging the images and eliminating image splicing lines, wherein the merging of the images is to eliminate redundant pixel information in the overlapped area and align the images to be spliced according to the image registration result; eliminating image splicing lines and performing weighted average fusion processing near the splicing lines; the weighted average weighting function can use a gradual-in gradual-out method, the complexity is low, the speed is high, and smooth transition of images can be realized in the overlapping area. The gradual-in gradual-out method is to calculate a weight according to the distance from a pixel to be fused to the boundary of a coincidence region, the weight is changed linearly, and a fusion formula is as follows:
Figure GDA0002264094360000051
in the formula (d)1、d2Representing the weight of the pixel point (x, y) on the image corresponding to the overlapped part, and meeting the following conditions: d1+d2＝1,0＜d1,d2＜1；
d1、d2The calculation formula of (2) is as follows:
Figure GDA0002264094360000052
wherein xiIs the abscissa, x, of the pixel point to be fusedi、xrRespectively are the horizontal coordinates of the left and right boundaries of the image overlapping area;
the method comprises the steps of firstly constructing a multi-scale space, obtaining an optimal corner detection threshold value by utilizing a significance analysis model, extracting feature points, describing the feature points by utilizing an ORB descriptor, and finally realizing rapid matching by utilizing a Hamming distance and RANSAC method.
Drawings
FIG. 1 shows an algorithm flow diagram of the present invention;
FIG. 2 is a graph showing the gradual-in and gradual-out weight variation;
Detailed Description
The invention is further illustrated with reference to the following figures and detailed description.
With reference to fig. 1, first, it is understood what is called the ORB algorithm, which combines the FAST feature point detection method with the BRIEF feature descriptor, and performs improvement and optimization, and the present invention mainly introduces two parts, feature point detection and feature point description, of the algorithm.
1) Feature point detection
The ORB algorithm uses a gaussian pyramid structure and calculates its principal direction for each feature point, so that the detected feature points have scale invariance and rotation invariance.
(1) Firstly, establishing a scale space, constructing an image pyramid, and only one image is arranged on each layer, which is different from SIFT.
(2) Calculating the number n of feature points to be extracted of each layer according to a formula, detecting the feature points on images with different scales by using a FAST algorithm, sequencing according to FAST corner response values, reserving the first 2n points, then calculating Harris corner response values of the feature points, sequencing, and reserving the first n points as the feature points of the layer.
(3) The principal direction of the feature points is calculated. ORB proposes a gray centroid method, i.e. there is an offset between the gray of a corner point and the centroid in its neighborhood, and this vector is taken as the direction of the feature point.
Defining the moment of the neighborhood S of any one feature point p as:
Figure GDA0002264094360000071
where I (x, y) is the gray value at point (x, y).
The centroid of the neighborhood S is:
Figure GDA0002264094360000072
the included angle between the characteristic point and the mass center is defined as the main direction of the characteristic point: θ ═ arctan (M)0,1/M1,0)
2) Description of characteristic points
The ORB algorithm improves the BRIEF descriptor, namely, the rBRIEF description method enables the descriptor to have rotation invariance. The BRIEF descriptor is essentially a binary code string with the length of m, m point pairs are selected around the characteristic points, the gray value of each point pair is compared, and the descriptor is coded into a binary form.
A binary comparison criterion function τ is defined as:
Figure GDA0002264094360000073
in order to remove noise interference, an ORB algorithm selects an image block of 5 × 5 at a characteristic point, and replaces the gray value of the characteristic point with the average gray value of the image block after smoothing.
Selecting m point pairs near the feature points, and comparing to obtain a binary string with the length of m as a feature descriptor:
Figure GDA0002264094360000074
the ORB algorithm uses the above calculated principal directions of feature points to determine the direction of a feature descriptor in order to make the descriptor rotationally invariant. The m point pairs around the feature point are combined into a matrix S:
Figure GDA0002264094360000075
defining a rotation matrix corresponding to the characteristic point direction theta as RθCharacteristic point pair matrix S corresponding to direction thetaθ＝RθAnd S. Wherein the content of the first and second substances,
Figure GDA0002264094360000081
θ is the principal direction of the feature point.
The feature descriptors after determining the direction are: gm(p,θ)＝fm(p)|(xi,yi)∈Sθ
In order to improve the discrimination performance of the descriptors, the ORB uses greedy search to select 256 test point pairs with the largest variance and the lowest correlation from all possible binary tests to form the required feature descriptors.
On the basis, the invention discloses a real-time image splicing method for reconnaissance and evidence obtaining of an unmanned aerial vehicle for police, which comprises three steps of improved ORB algorithm, image registration and image fusion;
the improved ORB algorithm includes: the ORB is improved, a significant model based on space-frequency domain analysis is applied to the selection of the optimal threshold value in the characteristic extraction stage by combining with a KSW entropy method, and a multi-scale space is constructed by utilizing a Gaussian pyramid;
inputting an image I, firstly converting the image I from an RGB color space to a CIE L ab color space, wherein the image comprises three channels, namely a brightness channel (L channel) and two color channels (a channel and b channel) in a L ab color space, eliminating fine texture details in a time domain by adopting Gaussian blur for the color channels to obtain a feature map of the corresponding channel, calculating a feature map of a L channel by adopting a high-frequency enhanced Butterworth high-pass filter for the brightness channels, and finally combining the feature maps of the channels to form a saliency map of an original image;
any input image can be represented by a magnitude spectrum and a phase spectrum, wherein the phase spectrum contains texture detail information of the image, the magnitude spectrum contains light and shade contrast information, if only the phase spectrum is reserved, the obtained image significant features contain partial background information and interfere the detection of image feature points to cause a mismatching phenomenon, a high-pass filter can realize the purpose of sharpening the edge of a target and maximally reserve edge information by attenuating and inhibiting low-frequency components, and in order to enhance the image detail information, sharpen the edge of the target and reduce noise interference, a high-frequency enhanced Butterworth high-pass filter is adopted;
cut-off frequency of D0The nth order butterworth high pass filter is defined as:
Figure GDA0002264094360000091
wherein the content of the first and second substances,
Figure GDA0002264094360000092
representing the frequency midpoint (u, v) and the frequency rectangle center
Figure GDA0002264094360000093
The distance of (d);
inputting an image I, and combining feature maps of L, a and b to obtain a final significant feature SM as follows:
SM(x,y)＝||IL-IL(x,y)||+||Ia-Ia(x,y)||+||Ib-Ib(x,y)||
wherein, IL(x, y) is the corresponding characteristic image pixel value of the brightness channel of the original image after passing through a high-frequency enhanced Butterworth high-pass filter, Ia(x,y),Ib(x, y) are color channels, respectivelyCorresponding characteristic map pixel values, IL,Ia,IbRespectively, the average feature vectors of the corresponding channel images, | | | · | |, which is a two-norm Euclidean distance;
the selection of the threshold value should reasonably change along with the change of the image gray scale characteristics, and the selection of the optimal threshold value is determined by combining a KSW entropy method according to the image significant characteristics, and the specific steps are as follows:
let the image gray scale range be [0, L-1]The threshold t divides the data into A, B categories, and the corresponding probability distributions are { p } respectively0,p1,p2,...,pt},{pt+1,pt+2,p2,...,pL-1In which p isiFor the frequency of occurrence of the corresponding gray level, order
Figure GDA0002264094360000094
Then the entropy for A, B classes would be:
Figure GDA0002264094360000095
Figure GDA0002264094360000096
the total entropy of the image is H ═ HA+HB. The optimal threshold value T is
Figure GDA0002264094360000097
K is a proportionality coefficient, and since the threshold of the feature point is related to the pixel contrast of the image, the gray level difference of the image, namely the optimal threshold, is determined by the significant feature of the image and an entropy method;
extracting feature points by using a best threshold obtained by combining a Fast corner detection algorithm, and describing the feature points by using an rBRIEF descriptor for subsequent image registration;
the image registration comprises the steps of utilizing feature point matching and utilizing RANSAC algorithm to screen matching points;
the characteristic point matching is used for searching two characteristic points with the shortest distance in two groups of characteristic point sets by using a distance function, and the distance between two binary descriptors can be represented by using a Hamming distance, wherein the Hamming distance refers to the number of characters with different corresponding positions between two character strings with the same length. The smaller the Hamming distance is, the more similar the two binary descriptors are;
calculating the shortest Hamming distance and the next shortest Hamming distance of each feature point to obtain a group of feature point matching pairs, and considering that the two feature points are matched when the ratio of the shortest distance to the next shortest distance is less than a threshold value;
the method comprises the steps of selecting a certain number of samples at random to estimate model parameters by screening matching points by using a RANSAC algorithm, classifying the rest data according to the estimated parameters, determining that one part of data is within an allowable error range, namely an inner point, and removing an error matching point pair through multiple hypothesis verification if the part of data is an outer point;
the RANSAC algorithm needs to use a homography matrix H which describes the transformation relation between the coordinates of two image points, including translation, rotation, scaling and the like, and can find the position of a point in one image in the other image through the matrix H, and assume a pair of matching points p in an image 1 and an image 21(x,y)，p2The transformation relationship between (x ', y') is:
Figure GDA0002264094360000101
8 parameters of the matrix H can be calculated by 4 pairs of matching points, and the RANSAC algorithm comprises the following steps:
(1) setting an initial value of iteration times as 0, a maximum iteration time N, an internal point number threshold T1 and an error threshold T2;
(2) randomly selecting 4 pairs from a plurality of pairs of points to be matched, and calculating parameters of a transformation matrix H between two images;
(3) calculating the distance between the coordinate values of the other feature points after H transformation and the matching points of the feature points, if the distance is smaller than an error threshold value T2, considering the matching point pair as an inner point, and if the distance is not smaller than the error threshold value T2, calculating the number of the inner points;
(4) if the number of inliers is greater than the threshold number of inliers T1, the current model is saved as the optimal model. Otherwise, adding one to the iteration times, and turning to the step (2) to continue the next iteration;
(5) if the maximum iteration number N is reached, returning a group of interior points with the maximum number of corresponding interior points, and obtaining a transformation matrix H;
the image stitching means: after the images are registered, the images are spliced through image fusion, wherein the image fusion is the last step of image splicing and mainly comprises two parts: merging the images and eliminating image splicing lines, wherein the merging of the images is to eliminate redundant pixel information in the overlapped area and align the images to be spliced according to the image registration result; eliminating image splicing lines and performing weighted average fusion processing near the splicing lines; the weighted average weighting function can use a gradual-in gradual-out method, the complexity is low, the speed is high, and smooth transition of images can be realized in the overlapping area. The gradual-in gradual-out method is to calculate a weight according to the distance from a pixel to be fused to the boundary of a coincidence region, the weight is changed linearly, and a fusion formula is as follows:
Figure GDA0002264094360000111
in the formula (d)1、d2Representing the weight of the pixel point (x, y) on the image corresponding to the overlapped part, and meeting the following conditions: d1+d2＝1,0＜d1,d2＜1；
d1、d2The calculation formula of (2) is as follows:
Figure GDA0002264094360000112
wherein xiIs the abscissa, x, of the pixel point to be fusedi、xrRespectively, the abscissa of the left and right borders of the image overlap region.
As shown in fig. 2, d1 changes gradually from 1 to 0, and the corresponding d2 changes from 0 to 1, so that a smooth transition is achieved in the overlapping region of the images.
The method comprises the steps of firstly constructing a multi-scale space, obtaining an optimal corner detection threshold value by utilizing a significance analysis model, extracting feature points, describing the feature points by utilizing an ORB descriptor, and finally realizing rapid matching by utilizing a Hamming distance and RANSAC method.
Table 1 and table 2 show the comparison results of the registration rate and registration time of the improved ORB algorithm herein with BRISK algorithm, FAST-ORB algorithm, respectively. The FAST-ORB algorithm refers to an image feature point matching algorithm which is subjected to feature extraction by the FAST algorithm and is described by an ORB descriptor. As can be seen from table 4-1, the improved ORB algorithm herein has the advantage of high registration rate compared to other algorithms. Table 4-2 shows the comparison of the time used by each algorithm, and it can be seen that since the algorithm herein adds a step of significance analysis, the registration time is slightly higher than the other two algorithms, but is comparable, and the advantage of high speed is maintained. Based on tables 4-1 and 4-2, the experimental results show that the improved ORB algorithm maintains the superiority of speed, and the matching rate of the improved ORB algorithm is improved for images with changes of scale, rotation, visual angle, illumination and the like.
TABLE 1 image registration ratio comparison results
Figure GDA0002264094360000131
TABLE 2 image registration time comparison results
Figure GDA0002264094360000132

Claims (1)

1. The image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence collection is characterized by comprising three steps of improved ORB algorithm, image registration and image fusion;
the improved ORB algorithm includes: the ORB is improved, a significant model based on space-frequency domain analysis is applied to the selection of the optimal threshold value in the characteristic extraction stage by combining with a KSW entropy method, and a multi-scale space is constructed by utilizing a Gaussian pyramid;
inputting an image I, firstly converting the image I from an RGB color space to a CIE L ab color space, wherein the image comprises three channels of a L channel, an a channel and a b channel in a L ab color space, wherein a L channel is a brightness channel, the a channel and the b channel are color channels, and for the color channels, eliminating fine texture details by adopting Gaussian blur in a time domain to obtain a feature map of the corresponding channel;
any input image can be represented by a magnitude spectrum and a phase spectrum, wherein the phase spectrum contains texture detail information of the image, the magnitude spectrum contains light and shade contrast information, if only the phase spectrum is reserved, the obtained image significant features contain partial background information and interfere the detection of image feature points to cause a mismatching phenomenon, a high-pass filter can realize the purpose of sharpening the edge of a target and maximally reserve edge information by attenuating and inhibiting low-frequency components, and in order to enhance the image detail information, sharpen the edge of the target and reduce noise interference, a high-frequency enhanced Butterworth high-pass filter is adopted;
cut-off frequency of D0The nth order butterworth high pass filter is defined as:
Figure FDA0002495680140000011
wherein the content of the first and second substances,
Figure FDA0002495680140000012
representing the frequency midpoint (u, v) and the frequency rectangle center
Figure FDA0002495680140000013
The distance of (d);
inputting an image I, and combining feature maps of L, a and b to obtain a final significant feature SM as follows:
SM(x,y)＝||IL-IL(x,y)||+||Ia-Ia(x,y)||+||Ib-Ib(x,y)||
wherein, IL(x, y) is the corresponding characteristic image pixel value of the brightness channel of the original image after passing through a high-frequency enhanced Butterworth high-pass filter, Ia(x,y),Ib(x, y) are respectively the corresponding characteristic map pixel values of the color channels, IL,Ia,IbRespectively, the average feature vectors of the corresponding channel images, | | | · | |, which is a two-norm Euclidean distance;
the selection of the threshold value should reasonably change along with the change of the image gray scale characteristics, and the selection of the optimal threshold value is determined by combining a KSW entropy method according to the image significant characteristics, and the specific steps are as follows:
let the image gray scale range be [0, G-1]The threshold t divides the gray scale into A, B types, and the corresponding probability distributions are { p }0，p1，p2，...，pt}，{pt+1，pt+2，pt+3，...，pG-1In which p isiFor the frequency of occurrence of the corresponding gray level, i is a positive integer, order
Figure FDA0002495680140000021
Then the entropy for A, B classes would be:
Figure FDA0002495680140000022
Figure FDA0002495680140000023
the total entropy of the image is H ═ HA+HBThen the optimum threshold value T is
Figure FDA0002495680140000024
K is a proportionality coefficient, and since the threshold of the feature point is related to the pixel contrast of the image, the gray level difference of the image, namely the optimal threshold, is determined by the significant feature of the image and an entropy method;
extracting feature points by using a best threshold obtained by combining a Fast corner detection algorithm, and describing the feature points by using an rBRIEF descriptor for subsequent image registration;
the image registration comprises the steps of utilizing feature point matching and utilizing RANSAC algorithm to screen matching points;
the characteristic point matching is used for searching two characteristic points with the shortest distance in two groups of characteristic point sets by using a distance function, and the distance between two binary descriptors can be represented by using a Hamming distance, wherein the Hamming distance refers to the number of characters with different corresponding positions between two character strings with the same length. The smaller the Hamming distance is, the more similar the two binary descriptors are;
calculating the shortest Hamming distance and the next shortest Hamming distance of each feature point to obtain a group of feature point matching pairs, and considering that the two feature points are matched when the ratio of the shortest distance to the next shortest distance is less than a threshold value;
the method comprises the steps of selecting a certain number of samples at random to estimate model parameters by screening matching points by using a RANSAC algorithm, classifying the rest data according to the estimated parameters, determining that one part of data is within an allowable error range, namely an inner point, and removing an error matching point pair through multiple hypothesis verification if the part of data is an outer point;
the RANSAC algorithm needs to use a homography matrix H which describes the transformation relation between the coordinates of two image points, including translation, rotation, scaling and the like, and can find the position of a point in one image in the other image through the matrix H, and assume a pair of matching points p in an image 1 and an image 21(x,y)，p2The transformation relationship between (x ', y') is:
Figure FDA0002495680140000031
in the formula
Figure FDA0002495680140000032
Inner characterFor the parameters of the matrix H, 8 parameters of the matrix H can be calculated by 4 pairs of matching points, and the RANSAC algorithm comprises the following steps:
(1) setting an initial value of iteration times as 0, a maximum iteration time N, an internal point number threshold T1 and an error threshold T2;
(2) randomly selecting 4 pairs from the n pairs of points to be matched, and calculating parameters of a transformation matrix H between the two images;
(3) calculating the distance between the coordinate values of the other feature points after H transformation and the matching points of the feature points, if the distance is smaller than an error threshold value T2, considering the matching point pair as an inner point, and if the distance is not smaller than the error threshold value T2, calculating the number of the inner points;
(4) if the number of inliers is greater than the threshold number of inliers T1, the current model is saved as the optimal model. Otherwise, adding one to the iteration times, and turning to the step (2) to continue the next iteration;
(5) if the maximum iteration number N is reached, returning a group of interior points with the maximum number of corresponding interior points, and obtaining a transformation matrix H;
the image fusion is the last step of image splicing and is mainly divided into two parts: merging the images and eliminating image splicing lines, wherein the merging of the images is to eliminate redundant pixel information in the overlapping area and align the images to be spliced according to the image registration result; eliminating image splicing lines and performing weighted average fusion processing near the splicing lines; the weighted average weight function can use a gradual-in gradual-out method, the complexity is low, the speed is high, the smooth transition of the image can be realized in the overlapping area, the gradual-in gradual-out method calculates the weight according to the distance from the pixel to be fused to the boundary of the overlapping area, the weight is linearly changed, and the fusion formula is as follows:
Figure FDA0002495680140000041
d1、d2representing the weight of the pixel point (x, y) on the image corresponding to the overlapped part, and meeting the following conditions: d1+d2＝1,0＜d1,d2＜1；
d1、d2The calculation formula of (2) is as follows:
Figure FDA0002495680140000042
wherein xiIs the abscissa, x, of the pixel point to be fusedi、xrRespectively, the abscissa of the left and right borders of the image overlap region.
CN201610954653.2A 2016-10-27 2016-10-27 Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining Active CN106683046B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610954653.2A CN106683046B (en) 2016-10-27 2016-10-27 Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610954653.2A CN106683046B (en) 2016-10-27 2016-10-27 Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining

Publications (2)

Publication Number Publication Date
CN106683046A CN106683046A (en) 2017-05-17
CN106683046B true CN106683046B (en) 2020-07-28

Family

ID=58840349

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610954653.2A Active CN106683046B (en) 2016-10-27 2016-10-27 Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining

Country Status (1)

Country Link
CN (1) CN106683046B (en)

Families Citing this family (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107464252A (en) * 2017-06-30 2017-12-12 南京航空航天大学 A kind of visible ray based on composite character and infrared heterologous image-recognizing method
CN107369170A (en) * 2017-07-04 2017-11-21 云南师范大学 Image registration treating method and apparatus
CN107490377A (en) * 2017-07-17 2017-12-19 五邑大学 Indoor map-free navigation system and navigation method
CN109919971B (en) * 2017-12-13 2021-07-20 北京金山云网络技术有限公司 Image processing method, image processing device, electronic equipment and computer readable storage medium
CN108319961B (en) * 2018-01-23 2022-03-25 西南科技大学 Image ROI rapid detection method based on local feature points
CN108961276B (en) * 2018-04-04 2020-09-25 国网智能科技股份有限公司 Distribution line inspection data automatic acquisition method and system based on visual servo
CN108805812A (en) * 2018-06-04 2018-11-13 东北林业大学 Multiple dimensioned constant ORB algorithms for image mosaic
CN108921848A (en) * 2018-09-29 2018-11-30 长安大学 Bridge Defect Detecting device and detection image joining method based on more mesh cameras
CN109543561B (en) * 2018-10-31 2020-09-18 北京航空航天大学 Method and device for detecting salient region of aerial video
CN109712071B (en) * 2018-12-14 2022-11-29 电子科技大学 Unmanned aerial vehicle image splicing and positioning method based on track constraint
CN111353933A (en) * 2018-12-20 2020-06-30 重庆金山医疗器械有限公司 Image splicing and fusing method and system
CN109801220B (en) * 2019-01-23 2023-03-28 北京工业大学 Method for solving mapping parameters in vehicle-mounted video splicing on line
CN110211363B (en) * 2019-04-12 2020-07-10 台州市浩峰电器有限公司 Intelligent household appliance switch platform
CN110132302A (en) * 2019-05-20 2019-08-16 中国科学院自动化研究所 Merge binocular vision speedometer localization method, the system of IMU information
CN114143517A (en) * 2021-10-26 2022-03-04 深圳华侨城卡乐技术有限公司 Fusion mask calculation method and system based on overlapping area and storage medium
CN114283065B (en) * 2021-12-28 2024-06-11 北京理工大学 ORB feature point matching system and method based on hardware acceleration

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104167003A (en) * 2014-08-29 2014-11-26 福州大学 Method for fast registering remote-sensing image
CN104751465A (en) * 2015-03-31 2015-07-01 中国科学技术大学 ORB (oriented brief) image feature registration method based on LK (Lucas-Kanade) optical flow constraint

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9177225B1 (en) * 2014-07-03 2015-11-03 Oim Squared Inc. Interactive content generation

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104167003A (en) * 2014-08-29 2014-11-26 福州大学 Method for fast registering remote-sensing image
CN104751465A (en) * 2015-03-31 2015-07-01 中国科学技术大学 ORB (oriented brief) image feature registration method based on LK (Lucas-Kanade) optical flow constraint

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Markless augmented reality registration algorithm based on ORB;Yuanyuan Du 等;《2014 12th International Conference on Signal Processing (ICSP)》;20141023;第1236-1240页 *
基于 ORB 和改进 RANSAC 算法的图像拼接技术;佘建国 等;《江苏科技大学学报：自然科学版》;20150430;第29卷(第2期);第164-169页 *

Also Published As

Publication number Publication date
CN106683046A (en) 2017-05-17

Similar Documents

Publication Publication Date Title
CN106683046B (en) Image real-time splicing method for police unmanned aerial vehicle reconnaissance and evidence obtaining
Zhang et al. Deep-IRTarget: An automatic target detector in infrared imagery using dual-domain feature extraction and allocation
Ashraf et al. Dogfight: Detecting drones from drones videos
CN107507172A (en) Merge the extra high voltage line insulator chain deep learning recognition methods of infrared visible ray
WO2018076138A1 (en) Target detection method and apparatus based on large-scale high-resolution hyper-spectral image
CN103020985B (en) A kind of video image conspicuousness detection method based on field-quantity analysis
CN109614936B (en) Layered identification method for remote sensing image airplane target
Abidin et al. Copy-move image forgery detection using deep learning methods: a review
CN105427350A (en) Color image replication tamper detection method based on local quaternion index moment
CN113159043A (en) Feature point matching method and system based on semantic information
CN114972423A (en) Aerial video moving target detection method and system
CN111241943B (en) Scene recognition and loopback detection method based on background target and triple loss
Tang et al. Deep saliency quality assessment network with joint metric
CN105023264A (en) Infrared image remarkable characteristic detection method combining objectivity and background property
Sun et al. IRDCLNet: Instance segmentation of ship images based on interference reduction and dynamic contour learning in foggy scenes
Liu et al. Multi-scale feature fusion UAV image object detection method based on dilated convolution and attention mechanism
CN111833249A (en) UAV image registration and splicing method based on bidirectional point characteristics
Xing et al. Image-enhanced YOLOv5 and deep sort underwater multi-moving target tracking method
CN111898589A (en) Unmanned aerial vehicle image rapid registration method based on GPU + feature recognition
CN104573703A (en) Method for quickly identifying power transmission line based on partial derivative distribution and boundary strategy
CN106603888A (en) Image color extraction processing structure
Piao et al. Vdetor: an effective and efficient neural network for vehicle detection in aerial image
CN113793372A (en) Optimal registration method and system for different-source images
Cheng et al. BadPart: Unified Black-box Adversarial Patch Attacks against Pixel-wise Regression Tasks
CN111008555B (en) Unmanned aerial vehicle image small and weak target enhancement extraction method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
TR01 Transfer of patent right

Effective date of registration: 20220214

Address after: No.19 Keyuan Road, Lixia District, Jinan City, Shandong Province

Patentee after: Shandong public safety inspection and Testing Technology Co.,Ltd.

Address before: 250014 No. 19, ASTRI Road, Ji'nan, Shandong

Patentee before: INFORMATION Research Institute OF SHANDONG ACADEMY OF SCIENCES

TR01 Transfer of patent right