CN109840525B - Extraction and matching search method of circumferential binary features - Google Patents
Extraction and matching search method of circumferential binary features Download PDFInfo
- Publication number
- CN109840525B CN109840525B CN201711202580.2A CN201711202580A CN109840525B CN 109840525 B CN109840525 B CN 109840525B CN 201711202580 A CN201711202580 A CN 201711202580A CN 109840525 B CN109840525 B CN 109840525B
- Authority
- CN
- China
- Prior art keywords
- binary
- circumferential
- image
- algorithm
- bitmap
- 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
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Abstract
The invention relates to an extraction and matching search method of a circumferential binary feature, which comprises the steps of firstly, establishing a Gaussian pyramid simulation human eye imaging model by using a scale space, and detecting feature points of each layer of image of the Gaussian pyramid by using a FAST corner detection algorithm; calculating a circle point by using a center circle drawing algorithm; calculating the direction of the characteristic point by using a local image gravity center method; three rules of the mirror image invariance of the circumferential binary characteristic are provided, and a circumferential binary characteristic extraction algorithm is provided according to the rule of the mirror image invariance. After the binary features are calculated, a bitmap algorithm for quickly calculating the binary features is provided, and a bitmap corresponding to the binary features is calculated; and constructing a hash table by using the bitmap as a key, and optimizing the hash table query by using a bit set algorithm. The method solves the problem of low matching entry rate of the mirror image, has good mirror image invariance, and has higher entry rate and higher speed of the binary matching features searched by the bitmap locality sensitive hash algorithm.
Description
Technical Field
The invention relates to a method for extracting and matching and searching a circumferential binary feature, belonging to the field of image processing and machine vision.
Background
With the development of intelligent manufacturing, machine vision embodied as human intelligence is more important, and graphic image processing and vision are used as main methods for a robot to obtain external information, so that the automation and intelligence degree of industrial production can be improved. The binary characteristics have the advantages of fast calculation, effective matching and easy storage in image matching and positioning, can be used for three-dimensional reconstruction, medical modeling, image matching, target positioning and recognition, assisting machinery in avoiding obstacles and the like, and provides technical support for intelligent manufacturing and intelligent cities. The existing binary feature extraction algorithm has the problems of poor image invariance and low entry rate of a matched binary feature search algorithm.
In the binary feature extraction algorithm, the BRIEF features do not have mirror image invariance; the ORB features are relatively poor in vertical and horizontal mirror image invariance; the BRISK feature has vertical mirror image invariance and does not have horizontal mirror image invariance; the FREAK characteristic has poor invariance of horizontal and vertical mirror images; ORB, BRISK, FREAK all have horizontal vertical mirror invariance. Therefore, the entry rate of the binary feature matching source image and the mirror image is low, and the number of matching points is small. In the binary characteristic search algorithm, rotating densification bit arrangement hash, a deep hash algorithm for compact binary code learning, a deep learning algorithm for binary hash coding and minimum loss hash of compact binary codes are used for matching binary characteristic query; marius gives out a plurality of random KD tree algorithms to search for high-dimensional matching characteristics, a multi-level clustering tree algorithm is provided, in the high-dimensional data extensible neighbor algorithm, the approximate nearest neighbor algorithm is summarized, a corresponding function library for fast searching of the approximate nearest neighbor is given, and the searching algorithm has the problems of few matching points and long average searching time.
Disclosure of Invention
The invention aims to solve the technical problems that the existing binary feature extraction algorithm is poor in image invariance and low in input rate of a matched binary feature search algorithm. Aiming at the contents, a circle binary feature extraction and matching search algorithm is provided: when extracting the characteristics, simulating a human eye pinhole imaging model by using a Gaussian pyramid: the imaging of a near target is large, the imaging of a far target is small, the imaging of the near target is clear, and the imaging of the far target is fuzzy; detecting feature points by using a FAST algorithm; calculating the direction of the characteristic point by using an image brightness gravity center method; when the binary characteristics are searched and matched, a fast calculation bitmap algorithm and a bitmap locality sensitive hash algorithm are provided, and the entry rate of binary characteristic matching is increased.
The technical scheme adopted by the invention for realizing the purpose is as follows: the method for extracting and searching the matching binary features of the circumference comprises the following steps:
establishing a human eye imaging model: establishing a scale space and a Gaussian image pyramid by using a Gaussian fuzzy algorithm according to a set size proportion, wherein the scale space and the Gaussian image pyramid are used for simulating an eye imaging model;
detecting the characteristic points and determining the directions of the characteristic points: detecting the characteristic points of each image in the image pyramid; calculating a circumferential point by taking the characteristic point C detected by the FAST as a circle center, wherein the radius r of the circle belongs to [1, n ]; taking the characteristic point C as the center of a circle, and taking the brightness gravity centers of all pixel points within the maximum circle with the radius of n as the center of the circle to calculate the characteristic direction;
and (3) coordinate conversion: transferring the coordinates of all pixel points within the maximum circle to the direction of the characteristic point to obtain a rotating coordinate; calculating the interpolation gray scale of all the rotated circumferential points;
extracting the binary features of the circumference: for each circle with the radius of r, sequentially comparing the interpolation gray levels of the rotated circumferential points until the points on the circumference are traversed, calculating to obtain a circumferential binary character string, and performing mirror image invariant processing to integrate to obtain circumferential binary characteristics;
fast computation bitmap algorithm: rapidly calculating bitmaps corresponding to the circumferential binary features of the two images;
bitmap locality sensitive hashing algorithm: taking a bitmap or a part of bitmap of each binary characteristic of a source image as a key word, taking the key word and the ID of the binary characteristic as mapping, and storing the mapping into a hash table;
optimizing a query algorithm: performing optimization query on the keywords of the target image by using a bit set to match binary features in the hash table;
and (3) judging feature matching: and judging whether the binary characteristics corresponding to the keywords of the source image are matched with the binary characteristics corresponding to the ID mapped by the keywords of the target image according to the Hamming distance between the binary characteristics corresponding to the keywords of the source image and the binary characteristics corresponding to the ID mapped by the keywords of the target image.
The method for extracting the circle binary features comprises the following steps:
for each circle with radius r, sequentially comparing the interpolation gray levels of the rotated circumferential points until the points on the circumference are traversed, calculating to obtain a circumferential binary character string, carrying out mirror image invariant processing, and integrating to obtain circumferential binary characteristics
Circle point rotation to feature point square obtained by center circle drawing algorithmTo the calculation of the circular binary character string, starting from the position right above the circle center C, the circular points are numbered in turn clockwise along the circumference, and the interpolation gray scale of the corresponding points is I i The corresponding binary character is S i (ii) a When i is<At 11, S i Is calculated by the formula
When I is 11, adding I 11 And I 0 Comparison, i.e.
Mapping the source image binary string S to the level of the source image binary string S h Vertical mirror image binary string S of source image v And a horizontal and vertical mirror image binary string S of the source image t The reverse order is inverted, and the obtained character strings are respectively S' and S h ′、S′ v And S t ′；
S、S h 、S v And S t Respectively and respectively inversed character strings S' and S h ′、S′ v And S t ' bitwise OR, bitwise AND of the first half and the second half of the obtained character string to obtain a binary character string of a certain circle;
and sequentially connecting the character strings of all circles according to the radius to obtain the binary characteristics of the circles, and storing the binary characteristics in an unsigned character type.
The fast computation bitmap algorithm comprises the following steps:
firstly, 5 bits are selected from an unsigned character type number to form a 5-bit unsigned type number F i Then the length is 32 circumferential binary features of unsigned character type, resulting in 32F i (ii) a F is to be i Numbering 1 to 32 according to the sequence of the corresponding unsigned type number in the circumferential binary characteristics, and recording the corresponding record of each circumferential binary characteristic as F ═ F- 1 F 2 …F 32 ；F i Belong to [0,31 ]]；
Then, calculate the bitmap, record F i 0 to 31, a corresponding bit vector is set to B i Will be bit vector B i Converting into 32bit unsigned integer number, and storing in the memory;
and finally, taking the bit vector of the adjacent record as a bitmap of the circumferential binary characteristic according to the bit or the obtained target bit vector.
The bitmap locality sensitive hashing algorithm comprises the following steps:
taking a bitmap or partial bitmap of each binary characteristic of a source image as a key word, taking the key word and an ID (identity) of the binary characteristic as mapping, storing the mapping into a hash table, and constructing a plurality of hash tables; one hash bucket in each hash table corresponds to one key, and one key corresponds to one or more binary feature IDs.
The optimized query algorithm comprises the following steps: for the keyword key of the target image, a bit set bitset is constructed to perform optimization query on the matched binary features in the hash table:
the construction formula is as follows: bitset [ key/32] | (1 < (key% 32))
The query formula is: bitset [ key/32] & (1 < (key% 32))! Is equal to 0
If the current calculation result of the query formula is 0, the keyword does not exist, and the query is not carried out;
and if the calculation result of the query formula is not 0, the keyword exists, and then the bucket corresponding to the current hash table is queried.
The invention has the following advantages and effects:
1. the method can be used for mirror image matching and has strong adaptability. The method comprises the following steps of calculating the direction of the characteristic by using an image pyramid model, extracting by using a circumferential point, carrying out mirror image invariance treatment, and being used for mirror image, rotation image, image with distorted visual field and JEPG image matching and having mirror image invariance, rotation invariance, scale invariance and the like.
2. The efficiency of feature extraction and matching is high. The method comprises the steps of extracting features by using a binary character string form, achieving high speed, calculating a bitmap corresponding to the binary features by using an FCBM algorithm, providing a BMLSH algorithm to construct a search hash table, and optimizing and querying the hash table by using a bit set algorithm.
3. The matching effect is good. Experiments prove that compared with other algorithms, the algorithm has the advantages of more entry points, high entry rate and small average reprojection error.
4. The application range is wide. The method can be used for positioning and identifying a planar object, extracting three-dimensional reconstruction point cloud, splicing images and the like.
Drawings
FIG. 1 is a flow chart of point feature extraction;
FIG. 2 is a flow diagram of a bitmap locality-sensitive hashing algorithm;
FIG. 3 is a source image and three mirror images;
(a) is a source image, (b) is a horizontal mirror image, (c) is a vertical mirror image, and (d) is a horizontal and vertical mirror image;
FIG. 4 is a source map corresponding to a binary string and a mirror map corresponding to a string;
(a) the source image circumference binary character string, (b) the horizontal mirror image circumference binary character string, (c) the vertical mirror image circumference binary character string, and (d) the horizontal vertical mirror image circumference binary character string.
Detailed Description
The invention is further described in detail below with reference to the drawings and the embodiments.
The circle binary feature extraction and matching search method comprises the following steps:
establishing a human eye imaging model: establishing a scale space and a Gaussian image pyramid by using a Gaussian fuzzy algorithm according to a size ratio of 1.2:1, and simulating an eye imaging model: the imaging size of a near object is large, the imaging size of a far object is small, the imaging of the near object is clear, and the imaging of the far object is fuzzy.
Detecting the characteristic points: the characteristic points of each image in the image pyramid are detected by a FAST algorithm and a non-maximum suppression algorithm, because the FAST characteristic point detection algorithm has high efficiency, and because the idea is to use circumferential points, the method is consistent with the idea of extracting circumferential binary characteristics.
Determining a circumferential point: and (4) taking the characteristic point C detected by the FAST as the center of a circle, and calculating a circumference point by using a center point circle drawing algorithm, wherein the radius r of the circle belongs to [1,13 ].
Determining the direction of the characteristic points: first, the p + q moment m of the pixel brightness in the maximum circle is calculated pq And calculating a gravity center point G of the maximum circle area, wherein the angle between the gravity center point G and a straight line where the circle center C is located is the direction of the characteristic point, I represents the gray scale of the pixel (x, y), the (x, y) represents the coordinate under a coordinate system taking the point C as the origin of coordinates, x is the abscissa, and y is the ordinate.
θ＝atan(m 01 ,m 10 )
And (3) coordinate conversion: the 512 opposite coordinates are rotated to the direction of the characteristic point to obtain the rotation coordinates, which is a coordinate system with the characteristic point C as the origin and the point CG as the Y axis e ,y e )，x e And y e Is int type, the coordinate of point b is obtained by rotating point e counterclockwise by theta degrees and is (x) b ,y b )， x b And y b Is float type, the coordinate after the rotation of the circumferential point e is
Coordinate (x) of the point b b ,y b ) Turning to an image coordinate system with the characteristic point C as an original point and CG as a Y axis, and calculating the interpolation gray scale on the coordinate of the rotation point b by using a quadratic interpolation algorithm
Wherein
x and y are float types.
Extracting a circumferential binary characteristic character string: and for each circle with the radius of r, sequentially comparing the interpolation gray levels of the rotated circumferential points until the points on the circumference are traversed, calculating to obtain a circumferential binary character string, carrying out mirror image invariant processing, and integrating to obtain a circumferential binary characteristic.
Taking r as 2 as an example, if fig. 3 shows that the number of the circumferential points obtained by the center circle-drawing algorithm is 12, one cell of the table (a) in fig. 3 represents one pixel, the neighborhood of the feature point C is a block with the size of 5 × 5, and (a) in fig. 3 is a rotated coordinate, the character string calculation starts from the right above the center point C, (the number in the cell represents the point to obtain a gray value by secondary interpolation, i.e., a mean gray value), the circumferential points are numbered from 0 to 11 in sequence clockwise along the circumference, and the interpolated gray value of the corresponding point is I i Such as I 0 ＝15，I 1 ＝14，I 11 17, etc. the corresponding binary character is S i ，i∈[0,11]When i is<At 11, S i Is calculated by the formula
When I is 11, adding I 11 And I 0 Comparison, i.e.
The circumferential binary string S of (a) in fig. 3 is S ═ S 0 S 1 …S 11 In fig. 3, the value corresponding to the circumferential binary string calculated in (a) is S-101111101001; FIG. 3 (a) shows horizontal mirroring as shown in FIG. 3 (b), which is the calculated circumferential binary string S in FIG. 3 h ＝011010000010; vertical mirroring in FIG. 3 (a) occurs as shown in FIG. 3 (c), which calculates a circumferential binary string S v 000010011010; the horizontal vertical mirroring in FIG. 3 (a) occurs as shown in FIG. 3 (d), which results in a circumferential binary string S t 101001101111 before the circular binary string is characterized, a mirror invariant process is required.
The graphs shown in fig. 4 are circular binary string graphs corresponding to the graphs in fig. 3, respectively. A circumferential binary string S, S is presented herein h 、S v And S t The law between them; namely 3 relevant rules of image invariance of the circumferential binary feature:
the first half of the string S is S t The latter half of S is S t The front half of (2);
character string S h Is S in the first half of v The latter half of (1), S h The second half of (A) is S v The first half of (a);
the reverse order of the string S is S h 、S t The inverse of the order is S v .
Respectively to character strings S, S h 、S v And S t The reverse order is inverted, and the obtained character strings are respectively S' and S h ′、S′ v And S t '. according to the correlation law between circumferential binary strings: and bitwise combining the source character string and the inverse character string thereof or bitwise combining the first half part and the second half part of the obtained character string to obtain the final character string with mirror image invariability.
For example, the following steps are carried out: starting from a point right above the point C as in FIG. 4 (a), a character string S is obtained, S including S 0 To S 11 The reverse string of string S is S', in turn comprising S 1 ′ 1 To S 0 ', wherein S i Is' is S i Taking the inverse; the first half and the second half of the new character string are subjected to bitwise OR by the S and the S ', and the corresponding positions of the first half and the second half of the new character string are subjected to bitwise AND, as shown in the vertical form of example 1, the left side is the bitwise OR of the S and the S ', and the right side is the bitwise OR of the S and the S ', so that the first half and the second half of the new character string are subjected to bitwise OR; to S h 、S v And S t The same processing is performed, and as shown in the vertical forms of examples 2 to 4, the same binary character string 101011 is finally obtained.
Example 1
Example 2
Example 3
Example 4
After image invariant processing, the circumference binary character string obtained by calculating the source image and the 3 kinds of image images is the same character string, the processed circumference binary character string has image invariant, 13 circles are totally obtained from r to 1, the circumference binary character string obtained by calculating each circle is linked into a binary character string, and the binary character string is used as a circumference binary characteristic.
Fast computation bitmap algorithm: the bitmap index V is an aggregated bit vector of length N, each bit having a state that can only be 0 or 1. depending on the corresponding record, at bit i of the bit vector V, if the record at that position is t, the bit is 1, otherwise 0, for example: in the current file, there are 6 records, each record is a point pair of (int, binary), numbered 1 to 6, and sequentially: (30,101),(30,010),(40,011), (50,101),(40,010),(30,011).
In the example, there are 6 point pair records, so the bit vector length is 6, and the shaping records are 30 in the 1 st, 2 nd and 6 th, so the bit vector corresponding to the shaping record 30 is 110001; similarly, the bit vectors corresponding to the shaping records 40 and 50 are 001010 and 000100, respectively, in the example, the 1 st and 4 th binary records are 101, so the bit vector corresponding to the binary record 101 is 100100; similarly, the bit vectors corresponding to binary records 010 and 011 are 010010 and 001001, respectively.
Firstly, 5 bits are selected from a non-symbol character type number (8 bits) to form a 5-bit non-symbol type number F i If the length is 32 unsigned binary characteristics, 32F can be obtained by the above processing i A 1 to F i Numbering 1 to 32 in sequence according to the sequence of the corresponding unsigned type number in the binary characteristic, and then, corresponding record of each circumferential binary characteristic is F ═ F- 1 F 2 …F 32 ；F i Belong to [0,31 ]]I.e. recording F on each position of F i Belong to [0,31 ]](ii) a Then, a bitmap is calculated, and records 0 to 31 all have a corresponding bit vector set to B i I belongs to [0,31]Will be bit vector B i Converting into 32bit unsigned integer number, and storing in a memory; and finally, bit-wise or obtaining a final bit vector of the adjacent records as a bitmap of binary features.
TABLE 1 fast computation of bitmaps of circular binary features
Bitmap locality sensitive hashing algorithm: first, record F is calculated i : as shown in Table 1, U1 and U2 are data composed of the first four uchar of two neighboring binary features, M is a mask composed of four uchar for bit fetching, F i Is the final result of the bit fetch, i belongs to [1,4 ]]The ith uchar. fetching bit representing the binary feature is a new 5-bit number formed by erasing the state of the binary number corresponding to the uchar type of the feature at the position where the mask is 1 and erasing the other positions, and each uchar type of the mask M has only 5 bits of 1, as shown in Table 1, the first uchar type of the mask M is converted into the 1 st, 3 rd, 5 th, 7 th bit sum from the left of the binary numberSince the 8-th position is 1, F is obtained by taking the 1 st, 3 rd, 5 th, 7 th and 8 th positions from the left of U1 1 00000, converted to decimal 0; similarly, taking the 1 st, 3 rd, 5 th, 7 th and 8 th positions of U2 to obtain F 1 01000, 8 in decimal system; binary string F 1 00000 and F 1 ' -01000, the result of the bitwise exclusive-or is 01000, with one 1 in the result, so a binary string F 1 And F 1 ' Hamming distance between them is 1, binary string F 1 ' is F 1 1 neighbor of (c) and, similarly, F 2 ' equal to F 2 ，F 3 ' is F 3 Two neighbors of (F) 4 ' equal to F 4 From F 1 ' and F 1 、F 2 ' and F 2 As can be seen from the bit fetching process, the mask M mainly functions to avoid different bits in neighboring binary features and reduce the dimension of the binary feature vector when calculating the bitmap, as shown in table 2, in U1, Fi is 0101 for a bitmap of 12; in U2, F i The bitmap of' 27 is 0010.
TABLE 2 record F i And F i ' corresponding bit vector
Generally, a plurality of hash tables are built, wherein one hash bucket in the hash tables corresponds to one keyword, and one keyword corresponds to one or more binary characteristics ID., because the circumferential binary characteristics consist of 256 bits, namely 32 uchar types, the circumferential binary characteristic bitmap is a 32-dimensional bit vector and can be converted into 32-bit unsigned type numbers to be stored in an internal memory.
Optimizing a query algorithm: the bit set is an algorithm for quickly inquiring whether keywords exist, corresponding English is bitset, the text optimizes the inquiry of matched binary characteristics by using the bit set so as to quickly judge whether the current inquiry characteristics of the target image exist in the hash table, the initialized bit set is 0, firstly, all keywords key in the hash table are stored into the bit set according to a formula (4), when inquiring, whether the current inquiry keywords of the target image exist in the hash table is judged according to a formula (5), and the current calculation result is 0, so that the keywords do not exist; and when the calculation result is not 0, if the keyword exists, querying a bucket corresponding to the current hash table.
bitset[key/32]|＝(1＜＜(key％32)) (4)
bitset[key/32]&(1＜＜(key％32))！＝0 (5)
Storing query information and judging feature matching: in the bit set, if the current keyword of the target image exists in the hash table, inquiring the binary characteristics corresponding to all IDs in the hash bucket corresponding to the keyword, calculating the Hamming distance between the binary characteristics corresponding to the current keyword and the binary characteristics corresponding to the ID, and when the nearest neighbor characteristics and the next nearest neighbor characteristics of the current inquired binary characteristics are met, saving the corresponding ID and the distance between the nearest neighbor and the next nearest neighbor 1 The next nearest neighbor distance is d 2 Judging whether the current query feature is matched with the binary feature corresponding to the ID according to the NNDR algorithm, such as formula (6), if the threshold condition R is met<0.6, match, otherwise not match.
The invention is used for image feature extraction, and the feature extraction flow chart is shown in figure 1:
mixing the raw materials in a ratio of 1.2:1, establishing an image pyramid; detecting the characteristic points of each layer of the image pyramid by using a FAST detector non-maximum value inhibition method; calculating a circle point and a characteristic information extraction range by using a center circle drawing algorithm; calculating a circumferential binary string along each layer circumference; and carrying out mirror image invariance processing on the circumferential binary character string, and taking the processed binary character string as a circumferential binary characteristic.
The invention is used for image characteristic matching search, and the matching characteristic search flow chart is shown as figure 2:
firstly, calculating a bitmap: extracting binary features of the left image and the right image, and calculating bitmaps corresponding to the binary features in a segmented and sequential manner by using an FCBM algorithm; secondly, constructing a hash table: using a bitmap corresponding to the binary features of the left image as a key word, using the key word and the ID of the binary features as mapping, and storing the mapping into a corresponding bucket in each hash table, wherein one key word can correspond to the IDs of a plurality of binary features; thirdly, optimizing and searching the bit set: storing the keywords in the hash table into a bit set, because the bit set can quickly judge whether the current query keyword exists in the current hash table; fourthly, storing the query information: if the key word corresponding to the right image feature exists in the hash table, calculating the Hamming distance between the binary feature corresponding to the left image key word and the current right image feature, and storing the distance between the nearest neighbor and the next nearest neighbor of the current feature and the ID of the key word corresponding to the feature; fifthly, matching the entry points: judging whether binary features of the left image and the right image are matched by using an NNDR algorithm, calculating a rotation matrix of the left image to the right image point according to a left image matching point set and a right image matching point set, multiplying the rotation matrix by a distance between a coordinate obtained by the left image point and a coordinate corresponding to the right image point to be a projection error, judging whether a current point is an enclosure point according to the projection error, and dividing the enclosure point number by the matching point number to obtain a matching enclosure rate.
The main test environments are as follows:
operating the system: microsoft Windows7
Function library: OPENCV image processing function library
CPU：Pentium(R)Dual-Core
Dominant frequency: 2.93GHz
Memory: 2G
The embodiment takes a typical gallery as an example, and performs circle binary feature extraction and matching search.
Taking fig. 1 and 2 as an example, the specific process is as follows:
mixing the raw materials in a ratio of 1.2:1, establishing an image pyramid; detecting the characteristic points of each layer of the image pyramid by using a FAST detector non-maximum value inhibition method; calculating a circle point and a characteristic information extraction range by using a center circle drawing algorithm; calculating a circumferential binary string along each layer circumference; and carrying out mirror image invariance processing on the circumferential binary character string, and taking the processed binary character string as a circumferential binary characteristic.
Firstly, calculating a bitmap: extracting binary features of the left image and the right image, and calculating bitmaps corresponding to the binary features in a segmented and sequential manner by using an FCBM algorithm; secondly, constructing a hash table: using a bitmap corresponding to the binary features of the left image as a key word, using the key word and the ID of the binary features as mapping, and storing the mapping into a corresponding bucket in each hash table, wherein one key word can correspond to the IDs of a plurality of binary features; thirdly, bit set optimization search: storing the keywords in the hash table into a bit set, because the bit set can quickly judge whether the current query keyword exists in the current hash table; fourthly, storing the query information: if the key word corresponding to the right image feature exists in the hash table, calculating the Hamming distance between the binary feature corresponding to the left image key word and the current right image feature, and storing the distance between the nearest neighbor and the next nearest neighbor of the current feature and the ID of the key word corresponding to the feature; fifthly, matching the entry points: judging whether binary features of the left image and the right image are matched by using an NNDR algorithm, calculating a rotation matrix of the left image to the right image point according to a left image matching point set and a right image matching point set, multiplying the rotation matrix by a distance between a coordinate obtained by the left image point and a coordinate corresponding to the right image point to be a projection error, judging whether a current point is an enclosure point according to the projection error, and dividing the enclosure point number by the matching point number to obtain a matching enclosure rate.
The circular binary features can be well matched with mirror images, and have good adaptability to rotary images, scaling images and the like; according to the bitmap locality sensitive hash algorithm, the searched matched binary features are more, the time consumption is less, the entry rate is higher, the projection error is close to that of other algorithms, and the bitmap locality sensitive hash algorithm is suitable for search query of the matched binary features; the bitmap locality sensitive hashing is used for inquiring and matching the binary characteristics of the circumference, has high input rate and small average projection error, and can inquire and match mirror images. The algorithm can be used for the aspects of feature extraction and matching, image identification and positioning, time difference image acquisition in three-dimensional reconstruction, image splicing, robot navigation, non-contact measurement and the like.
Claims (5)
1. The method for extracting and matching the binary features of the circumference is characterized by comprising the following steps of:
establishing a human eye imaging model: establishing a scale space and a Gaussian image pyramid by using a Gaussian fuzzy algorithm according to a set size proportion, wherein the scale space and the Gaussian image pyramid are used for simulating an eye imaging model;
detecting the characteristic points and determining the directions of the characteristic points: detecting the characteristic points of each image in the image pyramid; calculating a circumferential point by taking the characteristic point C detected by the FAST as a circle center, wherein the radius r of the circle belongs to [1, n ]; using the feature point C as a circle center, and using the brightness centers of all pixel points within a maximum circle with the radius of n to calculate the feature direction;
and (3) coordinate conversion: transferring the coordinates of all pixel points within the maximum circle to the direction of the characteristic point to obtain a rotating coordinate; calculating the interpolation gray scale of all the rotated circumferential points;
extracting the binary features of the circumference: for each circle with the radius of r, sequentially comparing the interpolation gray levels of the rotated circumferential points until the points on the circumference are traversed, calculating to obtain a circumferential binary character string, carrying out mirror image invariant processing, and integrating to obtain a circumferential binary characteristic;
fast computation bitmap algorithm: rapidly calculating bitmaps corresponding to the circumferential binary features of the two images;
bitmap locality sensitive hashing algorithm: taking a bitmap or a part of bitmap of each binary characteristic of a source image as a key word, taking the key word and the ID of the binary characteristic as mapping, and storing the mapping into a hash table;
optimizing a query algorithm: performing optimization query on the keywords of the target image by using a bit set to match binary features in the hash table;
and (3) judging feature matching: and judging whether the binary characteristics corresponding to the keywords of the source image are matched with the binary characteristics corresponding to the ID mapped by the keywords of the target image according to the Hamming distance between the binary characteristics corresponding to the keywords of the source image and the binary characteristics corresponding to the ID mapped by the keywords of the target image.
2. The circumferential binary feature extracting and matching search method according to claim 1, wherein the extracting of the circumferential binary features includes the steps of:
for each circle with radius r, sequentially comparing the interpolation gray levels of the rotated circumferential points until the points on the circumference are traversed, calculating to obtain a circumferential binary character string, carrying out mirror image invariant processing, and integrating to obtain circumferential binary characteristics
The circle points obtained by the center circle drawing algorithm are rotated to the direction of the characteristic points, the calculation of the circle binary character string is started from the right upper part of the circle center C, the circle points are numbered in sequence along the circumference clockwise, and the interpolation gray scale of the corresponding points is I i Corresponding binary character being S i (ii) a When i is<At 11, S i Is calculated by the formula
When I is 11, adding I 11 And I 0 Comparison, i.e.
Mapping the source image binary string S to the level of the source image binary string S h Vertical mirror image binary string S of source image v And a horizontal and vertical mirror image binary string S of the source image t The reverse order is inverted, and the obtained character strings are S ' and S ' respectively ' h 、S′ v And S' t ；
S、S h 、S v And S t Respectively with respective reverse character strings S ', S' h 、S′ v And S' t Bitwise OR, bitwise AND of the first half and the second half of the obtained character string to obtain a binary character string of a certain circle;
and sequentially connecting the character strings of all circles according to the radius to obtain the binary characteristics of the circles, and storing the binary characteristics in an unsigned character type.
3. The circumferential binary feature extraction and matching search method according to claim 1, wherein the fast computation bitmap algorithm comprises the steps of:
firstly, 5 bits are selected from an unsigned character type number to form a 5-bit unsigned type number F i Then the length is 32 circumferential binary features of unsigned character type, resulting in 32F i (ii) a F is to be i Numbering 1 to 32 according to the sequence of the corresponding unsigned type number in the circumferential binary characteristics, and recording the record corresponding to each circumferential binary characteristic as F ═ F- 1 F 2 …F 32 ；F i Belong to [0,31 ]]；
Then, calculate the bitmap, record F i 0 to 31, a corresponding bit vector is set to B i Will be bit vector B i Converting into 32bit unsigned integer number, and storing in the memory;
and finally, bit vectors of adjacent records are subjected to bitwise or target bit vectors to serve as bitmaps of the circumferential binary features.
4. The circumferential binary feature extraction and matching search method according to claim 1, wherein the bitmap locality-sensitive hashing algorithm comprises the steps of:
taking a bitmap or partial bitmap of each binary characteristic of a source image as a key word, taking the key word and an ID (identity) of the binary characteristic as mapping, storing the mapping into a hash table, and constructing a plurality of hash tables; one hash bucket in each hash table corresponds to one key, and one key corresponds to one or more binary feature IDs.
5. The circumferential binary feature extraction and matching search method according to claim 1, wherein the optimized query algorithm comprises the following steps: for the keyword key of the target image, a bit set bitset is constructed to perform optimization query on the matched binary features in the hash table:
the construction formula is as follows: bitset [ key/32] | (1 < (key% 32))
The query formula is: bitset [ key/32] & (1 < (key% 32))! 0 ═ 0
If the current calculation result of the query formula is 0, the keyword does not exist, and the query is not carried out;
and if the calculation result of the query formula is not 0, the keyword exists, and then the bucket corresponding to the current hash table is queried.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711202580.2A CN109840525B (en) | 2017-11-27 | 2017-11-27 | Extraction and matching search method of circumferential binary features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711202580.2A CN109840525B (en) | 2017-11-27 | 2017-11-27 | Extraction and matching search method of circumferential binary features |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109840525A CN109840525A (en) | 2019-06-04 |
| CN109840525B true CN109840525B (en) | 2022-09-20 |
Family
ID=66879974
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711202580.2A Active CN109840525B (en) | 2017-11-27 | 2017-11-27 | Extraction and matching search method of circumferential binary features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109840525B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111813499A (en) * | 2020-07-08 | 2020-10-23 | 赛特斯信息科技股份有限公司 | Method for realizing policy verification aiming at container network and application system thereof |
| CN114494380A (en) * | 2022-04-18 | 2022-05-13 | 武汉大学 | Binary shape context feature descriptor construction method and point cloud registration method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103729654A (en) * | 2014-01-22 | 2014-04-16 | 青岛新比特电子科技有限公司 | Image matching retrieval system on account of improving Scale Invariant Feature Transform (SIFT) algorithm |
| CN104081435A (en) * | 2014-04-29 | 2014-10-01 | 中国科学院自动化研究所 | Image matching method based on cascading binary encoding |
| JP2015022383A (en) * | 2013-07-16 | 2015-02-02 | Kddi株式会社 | Searching apparatus, system, program, and method using binary local characteristic vector based on image |
| JP2015056077A (en) * | 2013-09-12 | 2015-03-23 | Kddi株式会社 | Image retrieval device, system, program, and method using image based binary feature vector |
| CN105373795A (en) * | 2015-09-18 | 2016-03-02 | 中国科学院计算技术研究所 | A binary image feature extraction method and system |
| CN106056601A (en) * | 2016-05-27 | 2016-10-26 | 北京理工大学 | Video-based human eye position detection method |
-
2017
- 2017-11-27 CN CN201711202580.2A patent/CN109840525B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2015022383A (en) * | 2013-07-16 | 2015-02-02 | Kddi株式会社 | Searching apparatus, system, program, and method using binary local characteristic vector based on image |
| JP2015056077A (en) * | 2013-09-12 | 2015-03-23 | Kddi株式会社 | Image retrieval device, system, program, and method using image based binary feature vector |
| CN103729654A (en) * | 2014-01-22 | 2014-04-16 | 青岛新比特电子科技有限公司 | Image matching retrieval system on account of improving Scale Invariant Feature Transform (SIFT) algorithm |
| CN104081435A (en) * | 2014-04-29 | 2014-10-01 | 中国科学院自动化研究所 | Image matching method based on cascading binary encoding |
| CN105373795A (en) * | 2015-09-18 | 2016-03-02 | 中国科学院计算技术研究所 | A binary image feature extraction method and system |
| CN106056601A (en) * | 2016-05-27 | 2016-10-26 | 北京理工大学 | Video-based human eye position detection method |
Non-Patent Citations (1)
| Title |
|---|
| 圆周二进制描述符的图像点特征提取方法;张展等;《计算机辅助设计与图形学学报》;20170815(第08期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109840525A (en) | 2019-06-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Vidanapathirana et al. | Locus: Lidar-based place recognition using spatiotemporal higher-order pooling | |
| Germain et al. | Sparse-to-dense hypercolumn matching for long-term visual localization | |
| Gao et al. | 3D model comparison using spatial structure circular descriptor | |
| CN111553409B (en) | Point cloud identification method based on voxel shape descriptor | |
| Tipaldi et al. | Geometrical flirt phrases for large scale place recognition in 2d range data | |
| CN102236794A (en) | Recognition and pose determination of 3D objects in 3D scenes | |
| CN109101981B (en) | Loop detection method based on global image stripe code in streetscape scene | |
| Drost et al. | Local hough transform for 3d primitive detection | |
| CN109840525B (en) | Extraction and matching search method of circumferential binary features | |
| CN115683129A (en) | Long-term repositioning method and device based on high-definition map | |
| Weinshall et al. | Stability and likelihood of views of three dimensional objects | |
| Cao et al. | Parallel K nearest neighbor matching for 3D reconstruction | |
| Ma et al. | 3D convolutional auto-encoder based multi-scale feature extraction for point cloud registration | |
| ElNaghy et al. | Taxonomy for 3D content-based object retrieval methods | |
| Yue et al. | Automatic vocabulary and graph verification for accurate loop closure detection | |
| Košnar et al. | Comparison of shape matching techniques for place recognition | |
| Lu et al. | Three-dimensional object recognition using an extensible local surface descriptor | |
| Wu et al. | A novel fusing semantic-and appearance-based descriptors for visual loop closure detection | |
| CN100357705C (en) | Method for recognising star map based on LPT transforms | |
| TW202244844A (en) | Image matching system | |
| Munoz et al. | Improving Place Recognition Using Dynamic Object Detection | |
| Hofstetter et al. | Reliable data association for feature-based vehicle localization using geometric hashing methods | |
| Fu et al. | Performance analysis of single-query 6-DoF camera pose estimation in self-driving setups | |
| Huishen et al. | An improved bag of words method for appearance based visual loop closure detection | |
| Qiao et al. | Visual Localization by Place Recognition Based on Multifeature (D-λLBP++ HOG) |
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 |