CN108256058B - Real-time response big media neighbor retrieval method based on micro-computing platform - Google Patents
Real-time response big media neighbor retrieval method based on micro-computing platform Download PDFInfo
- Publication number
- CN108256058B CN108256058B CN201810038892.2A CN201810038892A CN108256058B CN 108256058 B CN108256058 B CN 108256058B CN 201810038892 A CN201810038892 A CN 201810038892A CN 108256058 B CN108256058 B CN 108256058B
- Authority
- CN
- China
- Prior art keywords
- point
- points
- data points
- data
- binary
- 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.)
- Expired - Fee Related
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/43—Querying
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a real-time response big media neighbor retrieval method based on a micro-computing platform, which relates to the technical field of retrieval, and the technical scheme is to provide a novel method for quickly generating multimedia binary characteristics and a high-distinguishability post-verification distance table. Then, a Hash mapping function meeting the sequence preservation is learned, and the obtained binary code can be guaranteed to keep the original sequence relation between data points. And finally, establishing a high-distinguishability distance table according to the binary coding of the data points. The invention has the beneficial effects that: the problem of response to the big media data neighbor retrieval request in real time on a micro-computing platform is solved, the requirements on storage resources and computing power are low, and a neighbor retrieval system suitable for the micro-computing platform is constructed.
Description
Technical Field
The invention relates to the technical field of retrieval, in particular to a real-time response big media neighbor retrieval method based on a micro-computing platform.
Background
In recent years, the rapid development of multimedia technology has led to an exponential increase in shared multimedia data in networks, and people are beginning to pay attention to the problem of neighbor retrieval of massive multimedia data.
The conventional multimedia data is generally characterized by high-dimensional floating-point type vectors, for example, SIFT (Scale invariant feature transform) is characterized by 128-dimensional floating-point type vectors, while GIST is characterized by[2]Is a floating point type vector of 320 or 640 dimensions. If the multimedia neighbor points are searched by calculating and comparing the Euclidean distance between the floating point type characteristic vectors, the calculation complexity is higher, and the performance requirement on a calculation platform is more strict. Moreover, all floating-point data are all loaded into the memory of the computing platform at one time, and a large amount of storage resources are occupied. The two points make the micro-computing platform unable to respond to the near neighbor retrieval request of massive multimedia in real time. To this endIt has been proposed to convert high-dimensional floating-point type vectors into compact binary codes and to retrieve multimedia neighbors based on hamming distance relationships.
The earliest proposal for retrieving neighbor points by using binary coding is a locality sensitive hashing algorithm which generates a linear hashing mapping function satisfying locality sensitive characteristics in a random manner. However, the training process of the algorithm does not depend on the training data set, and the generated binary code should be long enough to obtain satisfactory neighbor search results.
In order to ensure that better neighbor search performance can be obtained by adopting compact binary codes, people begin to explore and utilize a machine learning method to generate binary codes meeting specific constraint conditions according to a training data set. Spectral hash algorithm[4]And constructing a spectrogram according to the similarity between the data points, and learning the binary codes of the data points by segmenting the similarity spectrogram. Iterative quantization hash algorithm[5]And the vertex of the hypercube is used as a coding center point, and the binary code of the data point is generated according to the distance relation between the data point and the vertex of the hypercube, but the vertex of the hypercube is fixed and unchanged, and the generated binary code has poor space distribution self-adaptive capacity. K-means hash algorithm[6]And the code center point is provided to meet the clustering distribution characteristic, so that the generated binary code is ensured to accord with the spatial distribution characteristic of the data set.
In order to obtain better neighbor search performance in hamming space, many hash algorithms establish similarity preserving constraints. Binary reconstruction hash algorithm[7]Requiring minimization of the difference between hamming and euclidean distances of a pair of data points, a kernel supervised hashing algorithm[8]The similarity of the data points in hamming space and euclidean space is required to have higher consistency. The definition in the hash algorithm described above belongs to the absolute similarity keeping constraint, but the approximate neighbor retrieval task focuses more on the relative similarity between data points. Ternary loss hash algorithm[9]The relative relation between triplets is required to be consistent in Hamming space and original space, and the sequence supervision Hash algorithm also defines an objective function based on the relative relation between the triplets tensor, and the topThe partial ordering supervised binary coding algorithm focuses on the relative similarity between the data points in the neighbor retrieval result which are ordered to be earlier.
Most hash algorithms adopt a random gradient descent algorithm to optimize an original objective function, but due to the limitations of the binary coding length and the random gradient descent algorithm, a single group of binary codes cannot accurately express the neighbor relation between data points. Therefore, the generation of multiple binary codes is explored, so that the problem of local optimal solution can be solved well. Charikar et al propose a multiple locality sensitive hashing algorithm (MLSH) that generates multiple binary codes for a data point in a random manner and returns near neighbors based on average hamming distances. K mean local sensitive Hash Algorithm (KLSH)[13]And generating a plurality of groups of initial central points in a random mode, and optimizing an objective function by using a random gradient descent algorithm to learn a plurality of groups of binary codes. Unlike the mechanism by which the KLSH algorithm generates each set of binary codes separately, the joint inverted index algorithm (JII) generates multiple binary code center points for data points at once, and then groups them according to their similarity relationship, thereby forming multiple sets of binary codes.
Disclosure of Invention
In order to realize the aim, the invention provides a real-time response big media neighbor retrieval method based on a micro-computing platform.
A real-time response big media neighbor retrieval method based on a microcomputer platform comprises the following steps:
a: learning global high-dimensional floating point feature vector set F ═ F of dataset to be queried1,F2,…,FnWhich in total comprises n eigenvectors.
b: in order to ensure higher consistency between the neighbor retrieval result in the Hamming space and the neighbor retrieval result in the original space, the invention defines a sequence retention constraint condition and a clustering distribution constraint condition.
The method comprises the following steps: the sequence holds the constraints.
In European space, according to Fm(1≤m≤n,FmRepresenting the mth high-dimensional floating-point vector in F) and FSorting the floating point vectors in the F according to the Euclidean distance between the complementary feature vectors, and obtaining the result as follows:after the floating point vector in F is mapped into binary code, another sort result can be obtained according to the Hamming distance relationship between the floating point vector and the binary code:the sequence preserving constraints in the present invention require that two different sequence numbers of data points are identical, i.e. have the same element at the same position in different sequences, which is defined as follows:
representing a feature vector, P, with sequence number m in Euclidean spacehAnd (c), returning the position sequence number of the feature vector in the Hamming space, wherein I (c) is a judgment function, and if the arrangement sequence numbers of the feature vector in the Hamming space and the Euclidean space are inconsistent, the value of the target function is increased. According to the invention, the target function is minimized, so that different serial numbers of the feature vectors have higher consistency, and a neighbor retrieval result with higher accuracy is obtained in a Hamming space.
Secondly, the step of: clustering distribution constraints.
If the floating point vector is mapped into binary code of length bit, then coexistence is 2bitA binary code. For a massive database, the number of data points is much greater than 2bitMultiple data points may be mapped to the same code. The invention requires that data points with the same binary code should accord with the cluster distribution characteristic, and the definition of the constraint condition is shown as the following formula:
C(Fm) Is represented by the formulamHave the same center point of binary encoding.
c: when finding a binary coding center point meeting the sequence preservation and clustering distribution constraints, a random gradient descent algorithm can be adopted to optimize R + A. At this time, in order to ensure that the algorithm can be converged to all minimum value points quickly, the minimum value points in the objective function can be roughly judged in advance, and an initial central point is selected near the minimum value points. Since the definition of A is similar to the constraint condition of the clustering algorithm, the clustering center point can be used for estimating the minimum value point of the objective function R + A.
The method comprises the following steps: uniformly sampling the area where the data points are distributed, and then counting the number of high-dimensional floating point features falling in each sampling area, and taking the number as the density value of the area.
Secondly, the step of: a large number of data points are typically clustered around the cluster center point, which has a large density value. If the density value of a certain region is lower than the average density value, the probability value of the region containing the cluster center point is smaller, and the region is abandoned.
③: if the distance between two regions is small, then selecting an initial point from the two regions causes the algorithm to converge to the same extreme point. Thus, the present invention will incorporate high density regions that are relatively close in distance.
Fourthly, the method comprises the following steps: after processing of the second step and the third step, taking the average value of all data points in the area as a candidate initial center point set E' ═ P1,P2,…,PtAnd f, wherein t candidate initial center points are included in total.
Fifthly: if t>2bit(bit represents the length of binary code, 2)bitRepresenting the number of encoded center points), step d is performed. Otherwise, the number of the central points to be searched is not less than the number of the minimum value points possibly existing in the target function, and only 2 needs to be randomly selected from the data setbitAnd (E) forming an initial central point set E by the t data points and the points in the E', and searching an optimal coding central point in the step E.
d: and constructing a plurality of groups of initial center point sets.
t>2bitIt means that the number of center points to be found is less than the number of minimum points in the objective function. If only one set of center points is used, all the minimum points in the objective function cannot be found. For this case, the present invention will construct sets of initial center points.
The method comprises the following steps: i denotes the serial number of the data point, j denotes the serial number of the initial center point set, and both have initial values of 1.
Secondly, the step of: initializing empty set EjAnd E'.
③: copy E' to E ".
Fourthly, the method comprises the following steps: put the ith data point in E' into EjAnd delete the ith data point from E ".
Fifthly: neutralizing E' with EjMinimum distance value delta between data pointskMaximum Point addition EjAnd delete it from E ".
δkIs defined as follows:
δkreturn to E' injOf all data points in (c), dkuDenotes the kth data point in E' with EjThe distance value between the u-th data points.
Sixthly, the method comprises the following steps: continuously repeating the step (v) until step (E)jIn which contains 2bitA data point, and EjAs a set of initial center points.
Seventh, the method comprises the following steps: the values of i and j are both incremented by 1.
And (v): repeating the step (c), (c) and (j) until the values of i and j are t, and obtaining t sets of initial central points.
Ninthly: comparing the obtained t groups of initial central point sets, merging the same sets, and finally obtaining l groups of initial central point sets { E }1,E2,…,ElAnd (1 is less than or equal to l and less than or equal to t), and then turning to the step e.
e: and d, according to the initial central point set in the step c or d, searching a coding central point set which simultaneously meets the sequence retention constraint condition and the clustering distribution constraint condition by adopting a random gradient descent algorithm.
The method comprises the following steps: and randomly distributing binary codes with the lengths of bit to the data points in the initial central point set.
Secondly, the step of: the position of each center point is updated by minimizing the sequence preserving error and the clustering distribution error using a gradient descent algorithm.
③: the data point is assigned the same binary code as its nearest code center point.
Fourthly, the method comprises the following steps: and (4) continuously repeating the steps II and III until convergence, and obtaining the central point corresponding to each binary code. If the input is multiple sets of initial center points from d, multiple sets of code center points { C are generated1,C2,...,ClIn which there are l sets of code centroids, and each set contains 2bitA code center point.
f: and (c) establishing a high-distinguishability distance table according to the candidate initial central point set E' in the c.
The method comprises the following steps: calculating and comparing candidate initial center points P1,P2,…,PtAnd set { C1,C2,...,ClAnd assigning the same binary code as the code central point closest to the candidate initial central point by using the distance value between the central points. The binary code set of the candidate initial center point is { B }1,B2,…,BtIn which B isi={bi1,bi2,…,bilDenotes PiI is more than or equal to 1 and less than or equal to t, bij(1. ltoreq. j. ltoreq.l) represents PiAccording to the set of coded center points CjAnd generating a binary code.
Secondly, the step of: the value in each position of the distance table is calculated. If the position index of the distance table is (B, B'), then { B }1,B2,…,BtBetween the data points in the (b) data stream that constitute a binary coded pair (b, b')The distance value is stored in this location.
g: generating global floating point features F for querying multimedia data on a micro computing platformqAnd according to { C1,C2,...,ClF generationqMultiple sets of binary codes bq1,bq2,…,bql。
h: calculating FqAverage Hamming distance between the multiple groups of binary codes and the multiple groups of binary codes of the data points to be inquired, and arranging the data points in the database to be inquired according to the sequence from small to large. And if only a few data points share the same small Hamming distance value, taking the data points as a final neighbor retrieval result, and turning to the step j. Otherwise, returning the data points with smaller average Hamming distance value as alternative query results, and turning to step i.
i: and according to the later verification distance table, re-judging the similarity between the data points in the alternative query result and the query data points, and reordering the data points.
The method comprises the following steps: and f, obtaining a plurality of distance sets { D ] from the distance table generated in the step f by using the plurality of groups of binary codes of the query data points and the plurality of groups of binary codes of the data points in the alternative query results as indexes1,D2,…,Dl}。
Secondly, the step of: if x identical distance set indexes exist in the set, subtracting x-1 from the occurrence number of the distance values in the set, and only keeping the distance values which only appear 1 time.
③: and comparing elements in all distance sets, taking the same element value from the distance sets, and taking the average value of the elements as a basis for measuring the similarity between the data points.
Fourthly, the method comprises the following steps: and repeating the step I and the step II until similarity values between all data points in the alternative query result and the query data points are obtained from the distance table.
Fifthly: and (5) reordering the alternative query results according to the similarity values obtained in the process, and turning to step j.
j: and returning the multimedia data corresponding to the binary coding features ranked earlier in the query result as a final multimedia neighbor query result.
The technical scheme provided by the embodiment of the invention has the following beneficial effects: the invention provides an effective mechanism for learning multiple groups of binary codes, which can remarkably improve the neighbor retrieval performance in a Hamming space. The invention defines the sequence retention constraint condition, belongs to the relative similarity retention constraint condition and can ensure that the neighbor retrieval performance obtained in the Hamming space is better. Meanwhile, the invention also defines a cluster distribution constraint condition, which can ensure that the coding center point accords with the cluster distribution characteristic and the coding performance is better.
The invention provides a method for establishing an initial central point set, which can ensure that an algorithm can be converged to all extreme points in a target function at a higher speed. The method determines whether to generate a plurality of groups of binary codes or not by judging the quantity relationship between the minimum value point and the code central point in the target function, and evaluates the similarity between the data points by adopting the average Hamming distance, thereby ensuring that the neighbor retrieval result obtained in the Hamming space has higher consistency with the neighbor retrieval result in the Euclidean space.
The invention establishes the high-distinguishability posterior verification distance table, and can further improve the neighbor retrieval performance in the Hamming space by means of the distance table. Different from a general post-verification distance table, the method takes the estimated distance value between the clustering center points as the similarity value between different binary indexes, and the accuracy is higher.
The invention firstly generates a plurality of groups of binary codes, a plurality of groups of code central points and a post-verification distance table with high distinguishability of the multimedia data to be inquired in an online training stage. The on-line inquiry stage is divided into three steps: firstly, extracting high-dimensional floating point type characteristics of query multimedia data; then, converting the multi-group code center points obtained in the training stage into multi-group binary codes, calling the multi-group binary codes of the query data points and the data points to be queried into an internal memory of a micro-computing platform, and obtaining an initial neighbor retrieval result by computing and comparing average Hamming distances between the multi-group code center points; and finally, obtaining the similarity value between the data point pairs from the distance table according to the binary index of the data point pairs, and reordering the neighbor retrieval results, thereby further improving the neighbor retrieval performance.
The invention aims to establish a system for responding to the retrieval of mass multimedia neighbor in real time on a micro computing platform, the Hash algorithm provided by the invention can generate binary codes for keeping relative similarity, the binary codes better meet the essential requirement of approximate neighbor retrieval, and a high-distinguishability posterior verification distance table is established, thereby being beneficial to further improving the neighbor retrieval performance.
Because the computing power and the storage resource of the micro-computing platform are relatively limited, the traditional high-dimensional floating point type characteristic vector can not be adopted to respond to the near neighbor retrieval request of massive multimedia in real time. Aiming at the problems, the invention designs a novel Hash algorithm, maps the high-dimensional floating point vector into a compact binary code, and has higher storage compression ratio and less occupied storage resources. When the adjacent points are searched in the Hamming space, the hardware instruction XOR operation of the computer can be called to calculate the Hamming distance, the calculation complexity is low, and the calculation speed is high.
The invention discloses a hash algorithm capable of keeping relative similarity between high-dimensional floating point vectors, which can ensure that the neighbor retrieval results obtained in a Hamming space and the neighbor retrieval results in an Euclidean space have higher consistency. The method is not only very simple, but also has stronger theoretical background. According to the method, firstly, a candidate initial central point set is established according to the spatial distribution density of the data points, and the algorithm can be guaranteed to be converged to the minimum value point of the target function at a higher speed. Then, the code center points satisfying the sequence preservation constraint and the cluster distribution constraint are learned, which can ensure that the generated binary code can preserve the relative similarity between the data points and conform to the cluster distribution characteristics of the data set. And finally, establishing a posterior verification distance table with strong distinguishability according to the candidate initial central point set and the encoding central point, and further remarkably improving the neighbor retrieval performance.
Drawings
FIG. 1 is a flow chart of an embodiment of the present invention for generating two sets of two-bit binary codes.
Fig. 2 shows the comparison result of the recall rate of the multi-group binary coding algorithm of the present invention and the industry-preferred multi-group binary coding algorithm on the Cifar10 data set according to the embodiment of the present invention.
FIG. 3 shows the comparison of the recall ratio of the multi-block binary coding algorithm of the present invention with the industry-preferred multi-block binary coding algorithm on the GIST1M data set.
Figure 4 is a comparison of adaptive distance tables of the present invention and the performance of a distance table commonly used in the industry to promote neighbor search on a Cifar10 dataset.
FIG. 5 is a comparison of the adaptive distance table of the present invention and the performance of the neighbor search of the distance table commonly used in the industry on the GIST1M dataset.
FIG. 6 is an example of the retrieval of a single image on a Cifar10 dataset by an embodiment of the present invention and by industry preferred algorithms.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative of the invention and are not intended to be limiting.
Example 1
The invention provides a real-time response big media neighbor retrieval method based on a micro-computing platform.
Firstly, training process implementation steps
1. Extracting a global feature descriptor of the multimedia image information: GIST characteristics.
2. Learning a candidate initial central point set according to the spatial distribution density of the training data set: e' ═ P1,P2,…,Pt}。
3. According to the relation between the number of the candidate initial center points and the coding length, constructing a plurality of groups of initial center point sets: { E1,E2,…,El}。
4. According to the initial central point sets, adopting a random gradient descent algorithm, and passing throughAnd minimizing the values of the sequence preservation constraint condition and the clustering distribution constraint condition to learn a plurality of groups of coding center point sets: { C1,C2,...,Cl}。
5. According to the candidate initial central point set E' ═ P1,P2,…,PtAnd a set of encoded centroids { C }1,C2,...,ClAnd establishing a high-distinguishability post-verification distance table.
6. And storing the binary characteristics, the coding center point set and the post-verification distance table of the mass multimedia on a magnetic disk to finish the training process.
Second, the implementation step of the search process
1. The query input is obtained by the micro computing platform and GIST features of the input image are extracted.
2. And generating multiple groups of binary features of the query image according to the coding center point set.
3. And calculating the average Hamming distance between the query data set and the data set to be queried, and returning the data points which are ranked more front as alternative query results.
4. If a large number of data points have the same Hamming distance with the query data point, turning to step 5, otherwise, turning to step 6.
5. And (4) taking the binary codes of the query data point and the data point to be queried as indexes, acquiring the similarity value between the query data point and the data point to be queried from the post-verification distance table, and reordering the alternative query results.
6. And returning the multimedia corresponding to the binary feature with the top rank in the query result as a final query result. Hereinafter, an embodiment of the present invention will be described in detail
GIST feature: a high-dimensional floating point type vector belongs to a global feature descriptor and is in one-to-one correspondence with an image. Therefore, the neighbor search result of the GIST feature is the neighbor search result of the multimedia data.
2. Random gradient descent algorithm: the algorithm updates the parameter values along the gradient direction of the objective function, so that the algorithm can be ensured to be quickly converged to a minimum value, but the algorithm has local optimality, and the convergence speed and the convergence result of the algorithm are related to the selected initial point. If the initial point is near the extreme point, the convergence speed of the algorithm is high, and the initial point is selected near different extreme points, so that the algorithm converges at different extreme points.
3. And (3) Hash algorithm: a method of mapping high-dimensional floating-point vectors into a low-dimensional Hamming space can represent high-dimensional floating-point vectors as compact binary encodings. The general hash algorithm needs to satisfy the local sensitivity characteristic, that is, similar data points in the original space are mapped to similar binary codes with a high probability.
RCH: a set of initial center points is generated in a random mode, and a set of binary codes of the data points are generated by utilizing an algorithm system in the invention.
KRCH: a plurality of groups of coding center points are generated in a random mode, and a plurality of groups of binary codes of the data points are generated by utilizing an algorithm system in the invention.
6, MRCH: the algorithm system provided by the invention generates a plurality of groups of coding center points according to the spatial distribution density of a data set, and obtains a plurality of groups of binary codes of data points by optimizing a sequence consistency objective function.
Example 2:
data set: the GIST1M data set and the Cifar10 data set, which are popular, are often used to test the performance of neighbor search systems. The GIST1M data set contained 10 ten thousand training samples, 100 ten thousand data points to be queried, and 1 ten thousand query sample data points. The Cifar10 dataset contains 10 classes of image sets, each of which contains 5000 training images and 1000 query images. The invention adopts GIST algorithm to extract the global feature descriptors of the training image and the test image in the Cifar10 data set.
Evaluation indexes are as follows: the recall rate and accuracy of the general indicators that can reflect the performance of the search are used to compare the present invention with other superior methods in the industry.
Commercially preferred methods for comparison with the present invention include JII, KLSH, MLSH and LSH.
And analyzing the neighbor retrieval result of each algorithm, and calculating to obtain the recall rate curve and the accuracy value of each algorithm on different data sets.
The comparison results of the protocol examples are shown in fig. 2, fig. 3, fig. 4, fig. 5, and fig. 6.
Fig. 2 shows the comparison result of the recall rate of the multi-group binary coding algorithm of the present invention and the industry-preferred multi-group binary coding algorithm on the Cifar10 data set. The code length in each row is 32, 64 and 128 from top to bottom, respectively. From left to right, the number of binary code groups in each column is 4 and 8, respectively.
FIG. 3 shows the comparison of the recall ratio of the multi-bin binary coding algorithm of the present invention and the industry preferred multi-bin binary coding algorithm on the GIST1M data set. The code length in each row is 32, 64 and 128 from top to bottom, respectively. From left to right, the number of binary code groups in each column is 4 and 8, respectively.
Figure 4 is a comparison of the adaptive distance table of the present invention and the distance table commonly used in the industry to improve neighbor search performance on a Cifar10 dataset. The code length in each row is 32, 64 and 128 from top to bottom, respectively. From left to right, the number of binary code groups in each column is 4 and 8, respectively.
Figure 5 is a comparison of the adaptive distance table of the present invention and the distance table commonly used in the industry to improve neighbor search performance on GIST1M data sets. The code length in each row is 32, 64 and 128 from top to bottom, respectively. From left to right, the number of binary code groups in each column is 4 and 8, respectively.
Figure 6 is an example of the retrieval of a single image on a Cifar10 dataset by the present invention and industry preferred algorithms.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (4)
1. A real-time response big media neighbor retrieval method based on a microcomputer platform is characterized by comprising the following steps:
a: learning global high-dimensional floating point feature vector set F ═ F of dataset to be queried1,F2,…,Fn-n eigenvectors are included;
b: in order to ensure higher consistency between the neighbor retrieval result in the Hamming space and the neighbor retrieval result in the original space, a sequence retention constraint condition R and a clustering distribution constraint condition A are respectively defined;
c: when a random gradient descent algorithm is used for searching a binary coding center point which meets a sequence retention constraint condition and a spatial distribution relation constraint condition, in order to enable the algorithm to be quickly converged to all minimum value points, the minimum value points in a target function are roughly judged in advance, and the input of the algorithm is initialized nearby; because the definition of A is similar to the constraint condition of the clustering algorithm, the minimum value point of the target function R + A can be estimated according to the clustering center point;
c 01: uniformly sampling distribution areas of the data points, then counting the number of high-dimensional floating point features falling in each sampling area, and taking the number as the density value of the area;
c 02: a large number of data points are generally gathered near the center point, and the density value is large; if the density value of the region is lower than the average value, the probability that the region contains the clustering center point is smaller, and the region with the smaller density value is discarded;
c 03: if the distance between two regions is small, the initial point selected from the two regions will make the algorithm converge to the same extreme point; therefore, high density regions that are relatively close in distance will be merged;
c 04: after the c02 and c03 processing, the average of all data points in the region is used as the candidate initial center point set E' ═ { P ═ P1,P2,…,PtH, containing t candidate initial center points;
c 05: if t>2bit(bit represents the length of binary code, 2)bitRepresenting the number of the coding center points), turning to the step d; otherwise, the number of the central points to be searched is not less than the number of possible minimum value points in the target function, and only 2 needs to be randomly selected from the data setbit-t data points together with points in E' form an initial set of center points E, and a step of repeatinge, searching an optimal coding center point;
d: constructing a plurality of groups of initial center point sets;
t>2bitindicating that the number of the central points to be searched is less than the number of the minimum value points in the objective function; if only one set of central points is used, all minimum value points in the target function cannot be found; for the situation, constructing a plurality of groups of initial center point sets;
e: according to the initial central point set in the step c or d, a random gradient descent algorithm is adopted to search for a coding central point set which simultaneously meets a sequence retention constraint condition and a clustering distribution constraint condition;
f: establishing a distance table, and establishing the distance table according to the candidate initial central point set E' in the c04 to ensure high distinguishability;
f 01: calculating and comparing candidate initial center points P1,P2,…,PtAnd set { C1,C2,...,ClGiving the candidate initial center point and the closest code center point to the same binary code, wherein the binary code set of the candidate initial center point is { B }1,B2,…,BtIn which B isi={bi1,bi2,…,bilDenotes PiI is more than or equal to 1 and less than or equal to t, bij(1. ltoreq. j. ltoreq.l) represents PiAccording to the set of coded center points CjThe generated binary code;
f 02: calculate the value in each position of the distance table, if the position index of the distance table is (B, B'), then { B }1,B2,…,Bt-the distance value between data points that can constitute a binary coded pair (b, b') is stored in that location;
g: generating global floating point features F for querying multimedia data on a micro computing platformqAnd according to { C1,C2,...,ClF generationqMultiple sets of binary codes bq1,bq2,…,bql;
h: calculating FqOfAverage Hamming distances between the group of binary codes and a plurality of groups of binary codes of the data points to be queried are arranged according to a sequence from small to large; if only a few data points share the same small Hamming distance value, the data points are used as the final neighbor retrieval result, and step j is carried out; otherwise, returning a data point with a smaller average Hamming distance value as an alternative query result, and turning to step i;
i: according to the later verification distance table, the similarity between the data points in the alternative query result and the query data points is judged again, and the alternative query result and the query data points are reordered;
j: returning the multimedia data corresponding to the binary coding features with the higher ranking in the query result as a final multimedia neighbor query result;
in the step b, the step (c),
the sequence preserving constraint is defined as: in European space, according to Fm(1≤m≤n,FmRepresenting the mth high-dimensional floating point vector in F) and the rest of the feature vectors in F, sorting the floating point vectors in F to obtain the following results:after the floating point vector in F is mapped into binary code, another sort result can be obtained according to the Hamming distance relationship between the floating point vector and the binary code: the sequence preserving constraint requires a high correspondence between two different orderings, i.e. having the same element at the same position in the different orderings, which is defined as follows:
representing a feature vector, P, with sequence number m in Euclidean spacehThe sequence number of the position of the feature vector in the Hamming space is returned, I (-) is a judgment function, and if the sequence numbers of the feature vector in the Hamming space and the Euclidean space are not consistent, the value of the target function is increased; by minimizing the value of the objective function, the sequence numbers of the characteristic vectors in different spaces have consistency, so that a neighbor retrieval result with higher accuracy is obtained in a Hamming space;
the cluster distribution constraint is defined as:
if a floating point vector is mapped into a binary code of length bit, it co-exists at 2bitBinary coding; for a massive database, the number of data points is much greater than 2bitThere will be multiple data points mapped to the same code; data points with the same binary code should meet the cluster distribution characteristics, and the constraint conditions are defined as follows:
wherein C (F)m) Is represented by the formulamHave the same center point of binary encoding.
2. The real-time response big media neighbor searching method based on the micro-computing platform as claimed in claim 1, wherein in the step d, the method for constructing the plurality of sets of initial center points comprises:
d 01: i represents the serial number of the data point, j represents the serial number of the initial central point set, and the initial values of the i and the j are both 1;
d 02: initializing empty set EjAnd E ";
d 03: copy E' to E ";
d 04: put the ith data point in E' into EjAnd deleting the ith data from E ″Point;
d 05: neutralizing E' with EjMinimum distance δ between data pointskAdding E at the point with the largest valuejAnd deleting it from E';
δkis defined as follows:
δkreturn to E' injOf all data points in (c), dkuDenotes the kth data point in E' with EjThe distance value between the u-th data point;
d 06: step d05 is repeatedly executed until EjIn which contains 2bitA data point, and EjAs a set of initial center points;
d 07: adding 1 to both the values of i and j;
d 08: repeating the steps d02 to d07 until the values of i and j are t, and obtaining t groups of initial central point sets;
d 09: comparing the obtained t groups of initial central point sets, merging the same sets, and finally obtaining l groups of initial central point sets { E }1,E2,…,ElAnd (1 is less than or equal to l and less than or equal to t), and then turning to the step e.
3. The real-time response big media neighbor searching method based on the micro-computing platform as claimed in claim 1, wherein the specific implementation method of the step e is as follows:
e 01: randomly distributing binary codes with different lengths of bit for the data points in the initial central point set;
e 02: updating the position of each central point by minimizing sequence preserving errors and clustering distribution errors by adopting a gradient descent algorithm;
e 03: distributing binary codes which are the same as the code central points closest to the data points;
e 04: repeating steps e02 ande03, until convergence, obtaining the center point corresponding to each binary code, if the input is multiple groups of initial center points obtained from d, then generating multiple groups of code center point sets { C1,C2,...,ClIn which there are l sets of code centroids, and each set contains 2bitA code center point.
4. The real-time response big media neighbor searching method based on the micro-computing platform as claimed in claim 1, wherein the specific implementation method of the step i is as follows:
i 01: using the multiple groups of binary codes of the query data points and the multiple groups of binary codes of the data points in the alternative query results as indexes to obtain multiple groups of distance sets { D ] from the distance table generated in the step f1,D2,…,Dl};
i 02: if x distance sets with the same index exist in the i01, subtracting x-1 from the occurrence number of the distance values in the set, and only keeping the distance values which only occur 1 time;
i 03: comparing elements in all distance sets, extracting the same element value from the elements, and taking the average value of the elements as a basis for measuring the similarity between data points;
i 04: repeating the steps i01 to i03 until similarity values between all data points in the alternative result and the query data point are obtained from the distance table;
i 05: and reordering the alternative query results according to the similarity value obtained in the step i04, and turning to the step j.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810038892.2A CN108256058B (en) | 2018-01-16 | 2018-01-16 | Real-time response big media neighbor retrieval method based on micro-computing platform |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810038892.2A CN108256058B (en) | 2018-01-16 | 2018-01-16 | Real-time response big media neighbor retrieval method based on micro-computing platform |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108256058A CN108256058A (en) | 2018-07-06 |
CN108256058B true CN108256058B (en) | 2021-05-07 |
Family
ID=62740773
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810038892.2A Expired - Fee Related CN108256058B (en) | 2018-01-16 | 2018-01-16 | Real-time response big media neighbor retrieval method based on micro-computing platform |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108256058B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109299097B (en) * | 2018-09-27 | 2022-06-21 | 宁波大学 | Online high-dimensional data nearest neighbor query method based on Hash learning |
CN112100243B (en) * | 2020-09-15 | 2024-02-20 | 山东理工大学 | Abnormal aggregation detection method based on massive space-time data analysis |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105550368A (en) * | 2016-01-22 | 2016-05-04 | 浙江大学 | Approximate nearest neighbor searching method and system of high dimensional data |
CN107944046A (en) * | 2017-12-15 | 2018-04-20 | 清华大学 | Extensive high dimensional data method for quickly retrieving and system |
-
2018
- 2018-01-16 CN CN201810038892.2A patent/CN108256058B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105550368A (en) * | 2016-01-22 | 2016-05-04 | 浙江大学 | Approximate nearest neighbor searching method and system of high dimensional data |
CN107944046A (en) * | 2017-12-15 | 2018-04-20 | 清华大学 | Extensive high dimensional data method for quickly retrieving and system |
Non-Patent Citations (3)
Title |
---|
Research on the humanlike trajectories control of robots based on the k-nearest neighbors;Wang Lei;Liu Zhaowei;《2017 Chinese Automation Congress (CAC)》;20171231;第7746页-第7751页 * |
提升近邻检索性能的二值编码算法;王振;《吉林大学博士学位论文》;20170630;全文 * |
熵选择多重二进制编码;赵宏伟,王振等;《吉林大学学报(工学版)》;20170131;第218页-第226页 * |
Also Published As
Publication number | Publication date |
---|---|
CN108256058A (en) | 2018-07-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108920720B (en) | Large-scale image retrieval method based on depth hash and GPU acceleration | |
CN107085607B (en) | Image feature point matching method | |
Jégou et al. | On the burstiness of visual elements | |
Paulevé et al. | Locality sensitive hashing: A comparison of hash function types and querying mechanisms | |
CN103336795B (en) | Video index method based on multiple features | |
Schindler et al. | City-scale location recognition | |
TW410304B (en) | Multidimensional data clustering and dimension reduction for indexing and searching | |
CN113918753B (en) | Image retrieval method based on artificial intelligence and related equipment | |
CN103902704B (en) | Towards the multidimensional inverted index and quick retrieval of large-scale image visual signature | |
CN110059198A (en) | A kind of discrete Hash search method across modal data kept based on similitude | |
CN104199923B (en) | Large-scale image library searching method based on optimal K averages hash algorithm | |
CN104036012A (en) | Dictionary learning method, visual word bag characteristic extracting method and retrieval system | |
CN102799614B (en) | Image search method based on space symbiosis of visual words | |
JPWO2013129580A1 (en) | Approximate nearest neighbor search device, approximate nearest neighbor search method and program thereof | |
CN111177435B (en) | CBIR method based on improved PQ algorithm | |
CN108256058B (en) | Real-time response big media neighbor retrieval method based on micro-computing platform | |
CN105760875B (en) | The similar implementation method of differentiation binary picture feature based on random forests algorithm | |
CN110083732B (en) | Picture retrieval method and device and computer storage medium | |
CN109918529A (en) | A kind of image search method based on the quantization of tree-like cluster vectors | |
CN108536772B (en) | Image retrieval method based on multi-feature fusion and diffusion process reordering | |
CN113920303B (en) | Convolutional neural network based weak supervision type irrelevant image similarity retrieval system and control method thereof | |
CN113779287A (en) | Cross-domain multi-view target retrieval method and device based on multi-stage classifier network | |
CN114595350B (en) | Billion-level image quick searching method | |
CN108959650A (en) | Image search method based on symbiosis SURF feature | |
Liu et al. | VIEW GRAPH CONSTRUCTION FOR LARGE-SCALE UAV IMAGES: AN EVALUATION OF STATE-OF-THE-ART METHODS |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210507 Termination date: 20220116 |