WO2020098296A1 - 一种图像检索方法及装置 - Google Patents

一种图像检索方法及装置 Download PDF

Info

Publication number
WO2020098296A1
WO2020098296A1 PCT/CN2019/094919 CN2019094919W WO2020098296A1 WO 2020098296 A1 WO2020098296 A1 WO 2020098296A1 CN 2019094919 W CN2019094919 W CN 2019094919W WO 2020098296 A1 WO2020098296 A1 WO 2020098296A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
queried
feature vector
similar
dimensional space
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.)
Ceased
Application number
PCT/CN2019/094919
Other languages
English (en)
French (fr)
Inventor
兰银河
刘国宝
王晓云
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Unionpay Co Ltd
Original Assignee
China Unionpay Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Unionpay Co Ltd filed Critical China Unionpay Co Ltd
Publication of WO2020098296A1 publication Critical patent/WO2020098296A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/53—Querying

Definitions

  • the invention relates to the field of image technology, in particular to an image retrieval method and device.
  • the trained model is usually used to extract the features of the image, and the similarity judgment model is used to calculate the similarity between the features of the test image and the features of multiple sample images in the database.
  • the label of the sample image with the highest similarity is used as the label of the test image. Due to the increasing size of the gallery data, it is difficult to obtain enough label data to train the model, resulting in problems such as low recognition efficiency and difficulty in model training.
  • Embodiments of the present invention provide an image retrieval method and device to solve the technical problem of complicated and time-consuming image retrieval process in the prior art.
  • An embodiment of the present invention provides an image retrieval method.
  • the method includes:
  • Converting the feature vector of the image to be queried into a binary code according to a conversion rule the conversion rule is determined according to at least each stored image in the image library;
  • the query result of the image to be queried is determined according to the similarity between the feature vector of the similar image and the feature vector of the image to be queried.
  • the method before converting the feature vector of the image to be queried into a binary code according to a conversion rule, the method further includes:
  • a K-dimensional space is determined according to each reference image, and the variance of the feature vectors of the reference images in each projection direction of the K-dimensional space is the same and the largest; each reference image includes the stored images and the image to be queried; K is a positive integer;
  • K sphere hash mapping functions of the feature vectors of the reference images are solved; the K sphere hash mapping functions are used as the conversion rules.
  • the conversion rule is determined in the following manner:
  • a K-dimensional space is determined according to each reference image, and the variance of the projection distance of the feature vector of each reference image in the K-dimensional space is the largest and the same; each reference image is each stored image;
  • K sphere hash mapping functions of the feature vectors of the reference images are solved; the K sphere hash mapping functions are used as the conversion rules.
  • the determining the K-dimensional space according to each reference image includes:
  • the K-dimensional space is determined.
  • the comparing the binary code of the image to be queried with the binary code of each stored image in the image library to determine a similar image of the image to be queried includes:
  • the stored image whose Hamming distance is less than the first threshold is determined to be a similar image of the image to be queried;
  • the query result of the image to be queried includes:
  • An embodiment of the present invention provides an image retrieval device.
  • the device includes:
  • the acquisition module is used to determine the feature vector of the image to be queried
  • a processing module configured to convert the feature vector of the image to be queried into a binary code according to a conversion rule, the conversion rule is determined at least according to each stored image in the image library; the binary code of the image to be queried is Compare the binary codes of the stored images in the image library to determine the similar images of the image to be queried; for each similar image, according to the similarity between the feature vector of the similar image and the feature vector of the image to be queried, Determining the query result of the image to be queried.
  • processing module is further used to:
  • a K-dimensional space is determined according to each reference image, and the variance of the feature vectors of the reference images in each projection direction of the K-dimensional space is the same and the largest; each reference image includes the stored images and the image to be queried; In the K-dimensional space, K sphere hash mapping functions of the feature vectors of the reference images are solved; the K sphere hash mapping functions are used as the conversion rules.
  • the conversion rule is determined by determining the K-dimensional space according to each reference image, and the variance of the projection distance of the feature vector of each reference image in the K-dimensional space is the largest and the same;
  • Each reference image is the stored images; in the K-dimensional space, K sphere hash mapping functions of the feature vectors of the reference images are solved; the K sphere hash mapping functions are used as the conversion rules .
  • processing module is specifically used to:
  • processing module is specifically used to:
  • An embodiment of the present invention provides an image retrieval device, including at least one processor and at least one memory, wherein the storage unit stores a computer program, and when the program is executed by the processor, causes the processing
  • the device executes the steps of any method provided in the embodiments of the present invention.
  • An embodiment of the present invention provides a computer-readable medium that stores a computer program executable by an image retrieval device, and when the program runs on the image retrieval device, causes the image retrieval device to execute the embodiment of the invention The steps of any method provided.
  • the feature vector of the image to be queried may be converted into a binary code according to a conversion rule, and the binary code of the image to be queried and each of the image libraries Binary codes of stored images are compared to select similar images of the image to be queried; for each similar image, the similarity between the feature vector of the similar image and the feature vector of the image to be queried is determined The query result of the image to be queried.
  • the embodiment of the present invention screens out similar images by comparing the binary codes, which greatly reduces the images required for calculating the similarity; by filtering the similar images , And then calculate the similarity of the feature vectors to further improve the recall and accuracy of image retrieval.
  • the similar images are filtered through the binary code and the similarity is determined based on the similar images and the image to be queried, the storage space of the image feature vectors needed to retrieve the similar images is reduced, and the entire image retrieval process can be directly in memory Completed, there is no need to exchange data in the hard disk and memory, which improves retrieval efficiency.
  • FIG. 1 is a schematic flowchart of an image retrieval method according to an embodiment of the present invention
  • FIG. 2a is a schematic diagram of an image retrieval result provided by an embodiment of the present invention.
  • FIG. 2b is a schematic diagram of an image retrieval result provided by an embodiment of the present invention.
  • 2c is a schematic diagram of an image retrieval result provided by an embodiment of the present invention.
  • FIG. 3 is a schematic structural diagram of an image retrieval device according to an embodiment of the present invention.
  • FIG. 4 is a schematic structural diagram of an image retrieval device according to an embodiment of the present invention.
  • the commonly used image retrieval method is to adopt a machine learning method, first perform feature extraction and then perform similarity calculation to determine the retrieval result.
  • Current industry-related research focuses on offline training data models.
  • deep learning can effectively extract image feature vectors.
  • a convolutional neural network structure can include convolutional layers, pooling layers, and fully connected layers. And softmax layer.
  • the convolutional layer and the pooling layer can be combined to form multiple convolutional blocks to facilitate the extraction of image features layer by layer.
  • the neural network model needs to carry out training model parameters on the basis of labeled data. With the expansion of the image database, it is difficult to obtain the required label data for the massive image data, which makes the model training data take a long time. Therefore, it is difficult to obtain accurate Search Results.
  • the audit agency when the audit agency is reviewing user qualifications, for example, the audit agency can obtain images such as ID cards and business licenses submitted by the merchant during the process of auditing the merchant's network access. During the audit process, it needs to submit to Shanghai The ID card, business license and other images review their compliance. The image can be compared as the image to be queried with the image in the merchant's material that has been put into the warehouse to verify whether the merchant is eligible for network access.
  • the audit process is mainly conducted by manual. With the increase of business, the volume of merchants is getting larger and larger. The volume of merchants that need to be audited can reach tens of millions, and manual audits will take longer and longer, leading to audits. low efficiency. And the space required to store the database is increasing.
  • embodiments of the present invention provide an image retrieval method, which is used to solve the problem of complicated and time-consuming image retrieval process in the prior art.
  • FIG. 1 is a schematic flowchart of an image retrieval method according to an embodiment of the present invention. As shown in FIG. 1, it specifically includes the following steps:
  • Step 101 Determine the feature vector of the image to be queried.
  • Step 102 Convert the feature vector of the image to be queried into a binary code according to a conversion rule, and the conversion rule is determined according to at least each stored image in the image library.
  • Step 103 Compare the binary code of the image to be queried with the binary code of each stored image in the image library to determine a similar image of the image to be queried;
  • Step 104 For each similar image, determine the query result of the image to be queried according to the similarity between the feature vector of the similar image and the feature vector of the image to be queried.
  • the feature vector of the image to be queried may be converted into a binary code according to a conversion rule, and the binary code of the image to be queried and each of the image libraries Binary codes of stored images are compared to select similar images of the image to be queried; for each similar image, the similarity between the feature vector of the similar image and the feature vector of the image to be queried is determined The query result of the image to be queried.
  • the embodiment of the present invention screens out similar images by comparing the binary codes, which greatly reduces the images required for calculating the similarity; by filtering the similar images , And then calculate the similarity of the feature vectors to further improve the recall and accuracy of image retrieval.
  • the similar images are filtered through the binary code and the similarity is determined based on the similar images and the image to be queried, the storage space of the image feature vectors needed to retrieve the similar images is reduced, and the entire image retrieval process can be directly in memory Completed, there is no need to exchange data in the hard disk and memory, which improves retrieval efficiency.
  • the image to be queried can be obtained, and there is no limit to the way of obtaining it.
  • the image to be queried can be obtained from the Internet through web crawling, or can be captured manually Obtain the image to be queried in a manner; further, in an actual implementation process, an image to be queried may be obtained according to actual conditions, or an image to be queried may also be obtained, which is not specifically limited.
  • image retrieval the image to be queried and the stored image in the image library can be characterized and stored in the form of a matrix of feature vectors.
  • the feature extraction of the image to be queried can use CNN and other convolutional neural network models to train large-scale data to extract image features; for the stored images in the image library, feature extraction can be performed in advance offline.
  • the feature extraction of the image to be queried is obtained through the convolutional neural network model in the image to be queried and the image retrieval model to be queried, and can be determined by multiple feature extraction layers.
  • step 102 there may be multiple ways to determine the conversion rule.
  • a possible implementation may be based on the sphere hash function, after obtaining the image to be queried, each of the stored images and the N images to be queried The feature vector is projected into the K-dimensional space, and then the hash function of K spheres is determined.
  • N feature vectors can be expressed as ⁇ x 1 , ..., x N ⁇ , where, Data matrix Of columns. D is the dimension of the feature vector. The data of the feature vector is centered on zero
  • K hash mapping functions H (x) ⁇ h 1 (x), ..., h K (x) ⁇
  • the expression of the h k function can be:
  • Is a quantization function Is a real-valued embedding function
  • g (x) [g 1 (x), ..., g K (x)] T.
  • each supersphere needs to contain the data of the first threshold, and the overlap of any two superspheres needs to contain the data of the second threshold to ensure that the data included in each supersphere Enough is enough, and each hypersphere is relatively independent.
  • the first threshold may be selected as 50%
  • the second threshold may be selected as 25%. Iteratively solve the hash mapping function according to these two conditions.
  • a threshold function is used to map the data of the feature vector to a binary code of 0 or 1.
  • the sphere hashing method is used, and the relatively non-linear hashing method has better spatial partitioning performance, so that the accuracy of the calculated similarity of the converted binary code is higher.
  • the method may further include:
  • K sphere hash mapping functions are determined according to each reference image; each reference image includes the stored images and the image to be queried; the K sphere hash mapping functions are used as the conversion rule.
  • the method further includes:
  • K sphere hash mapping functions are solved; the K sphere hash mapping functions are used as the conversion rules.
  • an embodiment of the present invention provides a conversion method, including:
  • Step 1 Determine the K-dimensional space according to each reference image, and the feature vector of each reference image has the largest variance in each projection direction of the K-dimensional space; each reference image includes the stored images and the image to be queried ;
  • Step 2 In the K-dimensional space, find K sphere hash mapping functions of the feature vectors of the reference images; the K sphere hash mapping functions serve as the conversion rule.
  • the PCA method of principal component analysis can be used to project the data into the K-dimensional space, where K can be equal to K, and K is the length of the binary code.
  • the dimension of the hypersphere is reduced, thereby reducing the problem of the accuracy of the similarity calculation of the binary code due to the increase of the dimension of the hypersphere.
  • M feature vectors of the stored images can also be projected into a K-dimensional space.
  • the N feature vectors The variance is the largest; in the K-dimensional space, solve the K sphere hash mapping functions of the N feature vectors; use the K sphere hash mapping functions as the conversion rules to determine the image to be queried K-bit binary code.
  • a possible implementation may determine the K-dimensional space according to each reference image, and the projection distance of the feature vector of each reference image in the K-dimensional space
  • the variance of is the largest; among them, the reference images are the stored images; for other implementation processes, reference may be made to the above-mentioned embodiments, which will not be repeated here.
  • the coverage of the hypersphere in each projection direction is the same, passing and ensuring that the data distribution in each projection direction is as consistent as possible can further improve the accuracy of calculating the similarity of the binary code.
  • it can include:
  • Step 1 Determine the K-dimensional space according to each reference image, and the variance of the feature vector of each reference image in each projection direction of the K-dimensional space is the same and the largest; each reference image includes the stored images and the target image Query image
  • Step 2 In the K-dimensional space, find K sphere hash mapping functions of the feature vectors of the reference images; the K sphere hash mapping functions serve as the conversion rule.
  • a possible implementation may determine the K-dimensional space according to each reference image, and the projection distance of the feature vector of each reference image in the K-dimensional space The variance of is the largest and the same; the reference images are the stored images; in the K-dimensional space, K sphere hash mapping functions of the feature vectors of the reference images are solved; the K spheres are scattered
  • the column mapping function serves as the conversion rule.
  • step one may include the following steps:
  • Step 1 Determine the projection matrix that projects the feature vector of each reference image onto the K-dimensional space, so as to maximize the variance of the projection distance of the feature vector of each reference image in each projection direction of the K-dimensional space;
  • Step 2 Rotate the projection matrix according to the orthogonal matrix, so that the variance of the projection distance of the feature vector of each reference image in each projection direction of the rotated K-dimensional space is the same;
  • Step 3 Determine the K-dimensional space according to the rotated projection matrix.
  • the isotropic projection space can be expressed as follows:
  • step one For the specific implementation process in step one, reference may be made to the above-mentioned method based on principal component analysis to generate a projection matrix that minimizes reconstruction errors, or the projection matrix may be obtained in other ways, which will not be repeated here.
  • an orthogonal matrix can be generated to rotate the PCA projection component so that the variance of each projection direction is isotropic. Since the orthogonal matrix Q does not change the sum of differences in different projection directions, the rotated projection matrix P can still minimize the reconstruction error.
  • the orthogonal matrix It is used to rotate the projection matrix W so that the variance of each component is equal.
  • Solving the orthogonal matrix Q such that the difference in the upper direction of the projection direction is equal (isotropic) is equivalent to finding the orthogonal matrix Q so that T and Z satisfy the following conditions:
  • the specific solution process may include:
  • Step 2 In each iteration, T k is generated by Z k-1 .
  • Q k is generated by the characteristic decomposition of T k as follows:
  • Step 3 Update Z k according to the following calculation method:
  • T k is generated by Z k-1
  • Z k is generated by T k
  • the orthogonal matrix Q is obtained by iteratively solving.
  • the final projection matrix P after rotation is:
  • An isotropic low-dimensional space is constructed by orthogonal rotation matrix. According to the characteristics of the hypersphere, the original data is mapped to an isotropic K-dimensional space, and then the sphere hash is applied in the K-dimensional space, which effectively reduces the dimension of the feature vector, and due to the isotropic variance, makes the super The construction accuracy of the sphere is higher, which can effectively improve the accuracy and recall rate of the similar images calculated by the binary code.
  • the K sphere hash function can be used to determine the K bits in the binary hash code of the image to be queried according to the function threshold, and the binary image stored in the image library K bits in the hash code.
  • the binary code of the stored image in the image library may be determined in advance before obtaining the image to be queried. After obtaining the image to be queried, the K-bit binary code of the image to be queried may be determined according to the predetermined K sphere hash mapping functions as the conversion rule.
  • the K-bit binary code of the image to be queried may be determined according to the predetermined K sphere hash mapping functions as the conversion rule.
  • a possible implementation manner can be determined by determining the Hamming distance between the binary code of the image to be queried and the binary code of each stored image, and determining the stored image whose Hamming distance is less than the first threshold Is a similar image of the image to be queried;
  • the distance between the binary codes can be measured by Hamming distance or asymmetric distance, which is not limited here.
  • a possible implementation manner may include:
  • the image library stores the feature vector of the image and the feature vector of the image to be queried as input data, and maps the feature vector to a binary code, which greatly reduces the storage space of the image and can Quickly determine similar images of the image to be queried.
  • the number of similar images and the number of mapped bits K are determined according to the actual image library, and are not limited herein.
  • the image library takes the GIST1M data set as an example.
  • the data set includes 1 million 960-dimensional GIST feature vectors.
  • Figures 2a, 2b, and 2c show the recall rates when the image library is mapped to 32bit, 64bit, and 128bit, respectively. Among them, iso-sph is the algorithm in the embodiment of the present invention. It can be seen that the image retrieval method provided in the embodiment of the present invention has a higher recall rate and better retrieval effect than in the prior art. And when the data point is mapped to 128bit, the recall rate of the first 5000 nearest neighbors of the image to be queried can reach 90%.
  • the time required to retrieve the first 5000 neighbors of an image to be queried is shown in Table 1. It can be seen that when the data is mapped to 128 bits, it takes only 9.70 milliseconds to retrieve the first 5000 similar images of an image to be queried in a million-level image library.
  • multiple similar images are selected in advance (for example, the number of similar images is less than 10000), and then the similarity measurement of the multiple similar images and the picture to be queried is performed to improve the subsequent The efficiency of the similarity calculation between the image to be queried and the stored image in the image library.
  • the step of determining the similar image of the image to be queried in the image retrieval method in the embodiment of the present invention can improve the review efficiency.
  • the similar images of the first 10 images to be queried are output according to sorting, and then the 10 similar images can be manually reviewed. If there is an image with the same image to be queried, the merchant is denied access to the network, and if there is no identical image, the merchant is permitted to access the network.
  • a possible implementation method can determine the similarity between the similar image and the image to be queried, and then give the audit result according to the set rules and image query results. For example, you can measure the similarity distance between the feature vector of the image to be queried and the feature vector of the output similar image, and set an audit threshold. When the distance between the feature vector of the image to be queried and the similar image is less than the audit threshold, it is considered If the similarity is too high, the merchant will be refused access to the network; otherwise, the merchant will be admitted to the network.
  • the result of the audit can be given immediately after confirming that the merchant submits the materials with a successful record, which further improves the audit efficiency.
  • a person skilled in the art may set a preset threshold according to experience and actual conditions, which is not specifically limited.
  • the embodiments of the present invention can also be applied to face recognition application scenarios, so as to realize rapid face recognition of a large number of image galleries.
  • similar images of the image to be queried can be determined according to the image retrieval method in the embodiment of the present invention, and then according to the feature vector of the similar image
  • the comparison of the similarity of the feature vectors to determine whether the image to be queried exists in the image library can realize rapid retrieval of face recognition, achieve online face recognition speed of millisecond level, and can greatly reduce the storage space of the image.
  • step numbers are only an exemplary representation of an execution process, and the sequence of the steps in this application is not specifically limited.
  • the present invention proposes an image retrieval method, which solves the problem of low efficiency of manually reviewing tens of millions of image data; by converting the feature vector of the image into a binary code, and by comparing the similarity between the binary codes, the effective neighbor image is determined, The number of effective neighbor images for similarity measurement is greatly reduced, and then the similarity calculation is performed according to the feature vectors of the neighboring images and the feature vectors of the image to be queried to obtain high-precision retrieval results, which effectively improves the retrieval speed of the images.
  • the storage space occupied by the image feature vector in the retrieval process can be greatly reduced, and the single-machine image retrieval can be realized.
  • the data does not need to be exchanged between the memory and the hard disk, that is, all the operation calculations that can be retrieved in the stand-alone memory.
  • the retrieval process there is no need to label the gallery, and it can be applied to the unlabeled gallery, which greatly reduces the cost of manual annotation and reduces the cost of model training.
  • FIG. 3 exemplarily shows a schematic structural diagram of an image retrieval device provided by an embodiment of the present invention.
  • the device includes an acquisition module 301 and a processing module 302.
  • An embodiment of the present invention provides an image retrieval device.
  • the device includes:
  • the obtaining module 301 is used to determine the feature vector of the image to be queried
  • the processing module 302 is configured to convert the feature vector of the image to be queried into a binary code according to a conversion rule, the conversion rule is determined according to at least each stored image in the image library; the binary code of the image to be queried and the Compare the binary codes of the stored images in the image library to determine similar images of the image to be queried; for each similar image, according to the similarity between the feature vector of the similar image and the feature vector of the image to be queried To determine the query result of the image to be queried.
  • processing module 302 is further used to:
  • a K-dimensional space is determined according to each reference image, and the variance of the feature vectors of the reference images in each projection direction of the K-dimensional space is the same and the largest; each reference image includes the stored images and the image to be queried; In the K-dimensional space, K sphere hash mapping functions of the feature vectors of the reference images are solved; the K sphere hash mapping functions are used as the conversion rules.
  • the conversion rule is determined by determining the K-dimensional space according to each reference image, and the variance of the projection distance of the feature vector of each reference image in the K-dimensional space is the largest and the same;
  • Each reference image is the stored images; in the K-dimensional space, K sphere hash mapping functions of the feature vectors of the reference images are solved; the K sphere hash mapping functions are used as the conversion rules .
  • processing module 302 is specifically used to:
  • processing module 302 is specifically used to:
  • an embodiment of the present application provides an image retrieval device. As shown in FIG. 4, it includes at least one memory 401 and a processor 402 connected to the at least one memory. Processing is not limited in this embodiment of the present application.
  • the specific connection medium between the processor 402 and the memory 401 is shown in FIG. 4 by using a bus connection as an example.
  • the bus can be divided into address bus, data bus, control bus and so on.
  • the computing device described in the embodiment of the present invention may further include an input device 403 and an output device 404.
  • the input device 403 may include a keyboard, a mouse, a touch screen, etc .
  • the output device 404 may include a display device, such as a liquid crystal display (Liquid Crystal Display, LCD), a cathode ray tube (Cathode Ray Tube, CRT), a touch screen, and the like.
  • the memory 401 stores instructions executable by at least one processor 402. By executing the instructions stored in the memory 401, the at least one processor 402 can perform the steps included in the aforementioned image retrieval method.
  • the processor 402 is the control center of the image retrieval device, and can use various interfaces and lines to connect various parts of the image retrieval device, by running or executing the instructions stored in the memory 401 and calling the data stored in the memory 401 To achieve image retrieval.
  • the processor 402 may include one or more processing units, and the processor 402 may integrate an application processor and a modem processor, where the application processor mainly processes an operating system, a user interface, and application programs, etc.
  • the modulation processor mainly handles wireless communication. It can be understood that, the foregoing modem processor may not be integrated into the processor 402.
  • the processor 402 and the memory 401 may be implemented on the same chip, and in some embodiments, they may also be implemented separately on separate chips.
  • the processor 402 may be a general-purpose processor, such as a central processing unit (CPU), a digital signal processor, an application specific integrated circuit (AppKication, Specific Integrated Circuit, ASIC), a field programmable gate array or other programmable logic device, a discrete gate, or a transistor
  • the logic device and the discrete hardware component can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present application.
  • the general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in conjunction with the embodiments of the present application may be directly embodied and executed by a hardware processor, or may be executed and completed by a combination of hardware and software modules in the processor.
  • the memory 401 is a non-volatile computer-readable storage medium, and can be used to store non-volatile software programs, non-volatile computer executable programs, and modules.
  • the memory 401 may include at least one type of storage medium, for example, may include flash memory, hard disk, multimedia card, card-type memory, random access memory (Random Access Memory, RAM), static random access memory (Static Random Access Memory, SRAM), Programmable read-only memory (ProgrammabKe Read OnKy Memory, PROM), read-only memory (Read OnKy Memory, ROM), live erasable programmable read-only memory (EKectricaKKy ErasabKe ProgrammabKe Read-OnKy Memory, EEPROM), magnetic memory, disk , CD, etc.
  • the memory 401 is any other medium that can be used to carry or store a desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.
  • the memory 401 in the embodiment of the present application may also be a circuit or any other device capable of realizing a storage function, which is used to store program instructions and / or data.
  • embodiments of the present application also provide a computer-readable medium that stores a computer program executable by an image retrieval device, and when the program runs on the image retrieval device, causes the image retrieval The device performs the steps of the image retrieval method.
  • the embodiments of the present invention may be provided as methods or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.
  • computer usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.
  • computer usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processing machine, or other programmable data processing device to produce a machine that enables the generation of instructions executed by the processor of the computer or other programmable data processing device
  • These computer program instructions may also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, the instructions The device implements the functions specified in one block or multiple blocks of the flowchart one flow or multiple flows and / or block diagrams.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operating steps are performed on the computer or other programmable device to produce computer-implemented processing, which is executed on the computer or other programmable device
  • the instructions provide steps for implementing the functions specified in one block or multiple blocks of the flowchart one flow or multiple flows and / or block diagrams.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (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)
  • Image Analysis (AREA)
  • Processing Or Creating Images (AREA)

Abstract

本发明公开了一种图像检索方法及装置。所述方法包括：确定待查询图像的特征向量；将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的；将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。

Description

一种图像检索方法及装置
相关申请的交叉引用
本申请要求在2018年11月15日提交中国专利局、申请号为201811358729.0、申请名称为“一种图像检索方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及图像技术领域，尤其涉及一种图像检索方法及装置。
背景技术
目前，由于现有的图库数量很大，在图像检索方面，现有的深度学习算法在单机对海量图像数据训练和检索时经常需要将数据进行内存和硬盘之间的交换，容易造成图像检索过程复杂且耗时长的问题。
另外，现有技术在进行图像检索时，通常是使用训练好的模型提取图像的特征，并使用相似度判断模型计算测试图像的特征和数据库中的多个样本图像的特征之间的相似度，并将相似度最大的样本图像的标签作为测试图像的标签。由于图库数据规模日益增大，很难获得足够的标签数据，进行模型的训练，导致识别效率很低，模型训练难度增加等问题。
发明内容
本发明实施例提供一种图像检索方法及装置，以解决现有技术中图像检索过程复杂且耗时长的技术问题。
本发明实施例提供一种图像检索方法，所述方法包括：
确定待查询图像的特征向量；
将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的；
将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；
针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。
一种可能的实现方式，所述将所述待查询图像的特征向量依据转换规则转换为二进制码之前，还包括：
根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间的各投影方向的方差相同且最大；所述各参照图像包括所述各存储图像和所述待查询图像；K为正整数；
在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种可能的实现方式，所述转换规则通过以下方式确定：
根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大且相同；所述各参照图像为所述各存储图像；
在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种可能的实现方式，所述根据各参照图像确定K维空间，包括：
确定将所述各参照图像的特征向量投影至K维空间的投影矩阵，以使所述各参照图像的特征向量在所述K维空间的各投影方向上的投影距离的方差最大；
根据正交矩阵旋转所述投影矩阵，以使所述各参照图像的特征向量在旋转后的K维空间的各投影方向上的投影距离的方差相同；
根据旋转后的投影矩阵，确定所述K维空间。
一种可能的实现方式，所述将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像，包括：
通过确定所述待查询图像的二进制码与所述各存储图像的二进制码的汉明距离，将汉明距离小于第一阈值的存储图像，确定为所述待查询图像的相 似图像；
所述针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果，包括：
根据所述待查询图像的特征向量与所述相似图像的特征向量的欧式距离，确定所述待查询图像的特征向量与所述相似图像的相似度距离；
若确定在所述相似图像中，存在相似度距离满足预设条件的相似图像，则确定存在与所述待查询图像相似的存储图像，并将所述满足预设条件的存储图像作为所述待查询图像的查询结果；否则，所述查询结果为所述待查询图像不存在相似的存储图像。
本发明实施例提供一种图像检索装置，所述装置包括：
获取模块，用于确定待查询图像的特征向量；
处理模块，用于将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的；将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。
一种可能的实现方式，所述处理模块，还用于：
根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间的各投影方向的方差相同且最大；所述各参照图像包括所述各存储图像和所述待查询图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种可能的实现方式，所述转换规则通过以下方式确定：根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大且相同；所述各参照图像为所述各存储图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种可能的实现方式，所述处理模块，具体用于：
确定将所述各参照图像的特征向量投影至K维空间的投影矩阵，以使所述各参照图像的特征向量在所述K维空间的各投影方向上的投影距离的方差最大；根据正交矩阵旋转所述投影矩阵，以使所述各参照图像的特征向量在旋转后的K维空间的各投影方向上的投影距离的方差相同；根据旋转后的投影矩阵，确定所述K维空间。
一种可能的实现方式，所述处理模块，具体用于：
通过确定所述待查询图像的二进制码与所述各存储图像的二进制码的汉明距离，将汉明距离小于第一阈值的存储图像，确定为所述待查询图像的相似图像；根据所述待查询图像的特征向量与所述相似图像的特征向量的欧式距离，确定所述待查询图像的特征向量与所述相似图像的相似度距离；若确定在所述相似图像中，存在相似度距离满足预设条件的相似图像，则确定存在与所述待查询图像相似的存储图像，并将所述满足预设条件的存储图像作为所述待查询图像的查询结果；否则，所述查询结果为所述待查询图像不存在相似的存储图像。
本发明实施例提供一种图像检索的设备，包括至少一个处理器、以及至少一个存储器，其中，所述存储单元存储有计算机程序，当所述程序被所述处理器执行时，使得所述处理器执行本发明实施例提供的任一方法的步骤。
本发明实施例提供一种计算机可读介质，其存储有可由图像检索的设备执行的计算机程序，当所述程序在图像检索的设备上运行时，使得所述图像检索的设备执行本发明实施例提供的任一方法的步骤。
本发明实施例中，在确定待查询图像的特征向量之后，可将所述待查询图像的特征向量依据转换规则转换为二进制码，通过所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，筛选出所述待查询图像的相似图像；针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。相比现有技术中直接与图像库中的图像进行相似度计算的方式，本发明实施例通过比较二进制码筛选出相似图像，极大的降低了计算相似度所需的图像；通过 筛选相似图像，再进行特征向量的相似度计算，进一步提高了图像检索的召回率和准确率。另外，由于通过二进制码筛选相似图像及根据相似图像与所述待查询图像确定相似度，都降低了检索相似图像所需存储的图像特征向量的存储空间，整个图像检索过程都可以在内存中直接完成，不需要进行硬盘和内存中的数据交换，提高了检索效率。
附图说明
为了更清楚地说明本发明实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简要介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域的普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例提供的一种图像检索方法的流程示意图；
图2a为本发明实施例提供的一种图像检索结果的示意图；
图2b为本发明实施例提供的一种图像检索结果的示意图；
图2c为本发明实施例提供的一种图像检索结果的示意图；
图3为本发明实施例提供的一种图像检索装置的结构示意图；
图4为本发明实施例提供的一种图像检索设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清楚，下面将结合附图对本发明作进一步地详细描述，显然，所描述的实施例仅仅是本发明一部份实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本发明保护的范围。
目前，常用的图像检索方法为采用机器学习方法，先进行特征提取再进行相似度的计算，确定检索结果。目前业界相关的研究集中在线下训练数据的模型，例如采用深度学习的方法能有效提取图像的特征向量，例如，一种卷 积神经网络的结构可以包括卷积层、池化层、全连接层和softmax层。其中，卷积层和池化层可以配合组成多个卷积块，以便于逐层提取图像特征。神经网络模型需要在有标签的数据的基础上进行训练模型参数，随着图像数据库的扩大，海量的图像数据难以获得所需的标签数据，使得模型的训练数据时间长，因此，难以获得准确的检索结果。
另外，在进行特征向量间的相似度计算时，需要将待查询图像的特征向量与大规模图像中的所有特征向量分别计算相似度距离，计算耗时很长，而且，利用深度学习进行图像检索，海量的图像数据的特征向量的数据量较大，需要巨大的存储空间来存储高维图像数据，因此，计算待查询图像与图像数据库中的图像的特征向量间的相似度时，需要在内存与硬盘进行数据交换，极大的降低了计算的速度，使得检索效率低。
在一种具体的场景中，审核机构在审核用户资格等场景中，例如审核机构审核商户入网过程中，可以获得商户提交的身份证、营业许可证等图像，在审核过程中，需要对上海提交的身份证、营业许可证等图像审核其合规性。所述图像可以作为待查询图像与已入库的商户材料中的图像进行比对，审核商户是否有入网资格。目前，审核过程中主要为人工进行审核，随着业务的增加，商户的体量越来越大，需要审核的商户体量可以达到千万级，人工审核耗时会越来越长，导致审核效率很低。并且存储数据库所需的空间日益增加。
基于此，为了节约人力成本、快速获得图像检索结果，本发明实施例提供一种图像检索方法，用于解决现有技术中图像检索过程复杂且耗时长的问题。
图1所示，为本发明实施例提供的一种图像检索方法对应的流程示意图，如图1所示，具体包括如下步骤：
步骤101，确定待查询图像的特征向量。
步骤102，将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的。
步骤103，将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；
步骤104，针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。
本发明实施例中，在确定待查询图像的特征向量之后，可将所述待查询图像的特征向量依据转换规则转换为二进制码，通过所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，筛选出所述待查询图像的相似图像；针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。相比现有技术中直接与图像库中的图像进行相似度计算的方式，本发明实施例通过比较二进制码筛选出相似图像，极大的降低了计算相似度所需的图像；通过筛选相似图像，再进行特征向量的相似度计算，进一步提高了图像检索的召回率和准确率。另外，由于通过二进制码筛选相似图像及根据相似图像与所述待查询图像确定相似度，都降低了检索相似图像所需存储的图像特征向量的存储空间，整个图像检索过程都可以在内存中直接完成，不需要进行硬盘和内存中的数据交换，提高了检索效率。
具体来说，步骤101之前，可以获取待查询图像，其获取方式并没有限制，以待查询图像为人脸图像为例，可以通过网络爬取从互联网中获取待查询图像，也可以通过人工拍摄的方式获取待查询图像；进一步地，实际实施过程中可以根据实际情况获取一张待查询图像，也可以获取待查询图像，具体不做限定。在图像检索中，待查询图像和图像库中的存储图像可以通过特征向量的矩阵的形式来表征和存储。
步骤101中，所述待查询图像的特征提取可以采用CNN等卷积神经网络模型训练大规模数据提取图像特征；针对图像库中的存储图像，可以在线下预先进行特征提取。具体的，对待查询图像的特征进行提取是通过待查询图像和待查询图像检索模型中的卷积神经网络模型得到的，可以通过多个特征提取层确定。
在步骤102中，所述转换规则的确定方式可以有多种方式，一种可能的实现方式，可以根据球体散列函数，在获得待查询图像后，将各存储图像和待查询图像的N个特征向量投影至K维空间，进而确定K个球体散列函数。
具体的，N个特征向量可以表示为{x 1,...,x N}，其中，
Figure PCTCN2019094919-appb-000001
为数据矩阵
Figure PCTCN2019094919-appb-000002
的列。D为特征向量的维度。特征向量的数据是以零为中心的
Figure PCTCN2019094919-appb-000003
要生成K位二进制码，则需要建立K个散列映射函数H(x)＝{h 1(x),...,h K(x)}，h k函数的表达式可以为：
h k(x)＝q k(g k(x))
其中，
Figure PCTCN2019094919-appb-000004
是实值嵌入函数，
Figure PCTCN2019094919-appb-000005
是量化函数。定义函数
Figure PCTCN2019094919-appb-000006
g(x)＝[g 1(x),...,g K(x)] T。
根据球体散列(SPH)等方法，设置每一个超球体都需要包含第一阈值的数据，并且任何两个超球体的重叠需要包含第二阈值的数据，以保证每个超球体中包括的数据足够多，且个超球体相对独立。例如，第一阈值可以选择为50％，第二阈值可以选为25％。根据这两个条件迭代求解散列映射函数。采用一个阈值函数将特征向量的数据映射为0或1的二进制码。
本发明实施例中，通过球体散列方法，相对非线性的散列方法在空间上分区的性能更好，使得转换后的二进制码计算的相似度的精度更高。
结合一种具体的场景中，例如，审核机构在审核用户资格等场景中，随着业务及商户的体量的增加，图像库中的存储图像需要及时更新，因此，可以采用线上的方法，实施本发明实施例中提出的图像检索方法。一种可能的实现方式，在确定待查询图像的特征向量后，在步骤102之前，所述方法还可以包括：
根据各参照图像确定K个球体散列映射函数；所述各参照图像包括所述各存储图像和所述待查询图像；所述K个球体散列映射函数作为所述转换规则。
在另一种可能的场景中，若确定图像库中存储的图像变动较小，可以预 先在线下先根据图像库的存储图像确定K个球体散列映射函数。具体的，在步骤101之前，所述方法还包括：
根据所述各存储图像求解K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
但球体散列在处理高维数据时，由于原始数据维度增大，每个超球体要保持覆盖同样多的数据点必须增加半径，这使得数据在空间中的分别变得稀疏，导致二进制码计算相似度的精度降低。
为进一步提高通过二进制码确定多种相似图像的准确度，本发明实施例提供一种转换方式，包括：
步骤一、根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间的各投影方向的方差最大；所述各参照图像包括所述各存储图像和所述待查询图像；
步骤二、在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
具体的，可以用主成分分析PCA方法将数据投影到K维空间，其中K可以与K相等，K为二进制码的长度。PCA得到的投影矩阵
Figure PCTCN2019094919-appb-000007
的行为协方差矩阵XX T的前K个最大的特征向量，λ＝[λ 1,...,λ K]为相应的最大特征值。每个特征值对应于沿着相应投影方向的方差。
通过将原始数据映射到一个K维数据空间中，使得超球体的维度降低，进而减少由于超球体维度增加，导致的二进制码计算相似度的准确性下降的问题。
基于图像库变化较小的场景，本发明实施例还可以将所述各存储图像的M个特征向量投影至K维空间，在所述K维空间的各投影方向，所述N个特征向量的方差最大；在所述K维空间中，求解所述N个特征向量的K个球体散列映射函数；将所述K个球体散列映射函数作为所述转换规则，确定所述待查询图像的K位二进制码。
针对线下预先确定K个球体散列映射函数中的场景，一种可能的实现方式，可以根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大；其中，所述各参照图像为所述各存储图像；其他实施过程，可以参照上述实施方式，在此不再赘述。
进一步的，由于超球体在每个投影方向上的覆盖范围是一样的，通过并保证数据在每个投影方向上的分布尽可能一致,还可以进一步提高二进制码计算相似度的准确性。具体的，可以包括：
步骤一、根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间的各投影方向的方差相同且最大；所述各参照图像包括所述各存储图像和所述待查询图像；
步骤二、在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
针对线下预先确定K个球体散列映射函数中的场景，一种可能的实现方式，可以根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大且相同；所述各参照图像为所述各存储图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种优选的实施例中，步骤一可以包括以下步骤：
步骤一、确定将所述各参照图像的特征向量投影至K维空间的投影矩阵，以使所述各参照图像的特征向量在所述K维空间的各投影方向上的投影距离的方差最大；
步骤二、根据正交矩阵旋转所述投影矩阵，以使所述各参照图像的特征向量在旋转后的K维空间的各投影方向上的投影距离的方差相同；
步骤三、根据旋转后的投影矩阵，确定所述K维空间。
其中，各向同性的投影空间可以表示为如下：
Figure PCTCN2019094919-appb-000008
s.t.[PXX TP T] 11＝...＝[PXX TP T] KK
其中
Figure PCTCN2019094919-appb-000009
是旋转后的投影矩阵。
在步骤一的具体实施过程可以参考上述根据主成分分析的方法，生成最小化重建误差的投影矩阵，也可以使用其他方式获得所述投影矩阵，在此不再赘述。
在步骤二中，可以生成一个正交矩阵来旋转PCA投影成分使其每个投影方向方差是各向同性的。由于正交矩阵Q不改变不同投影方向上方差的和，旋转的投影矩阵P仍然可以最小化重建误差。
具体的，正交矩阵
Figure PCTCN2019094919-appb-000010
用于旋转投影矩阵W,使得每个分量的方差相等。求解正交矩阵Q使投影方向上方差相等(各向同性)相当于找到正交矩阵Q，使得T、Z满足以下条件：
||T-Z|| F＝0
其中T∈τ(a)，
Figure PCTCN2019094919-appb-000011
diag(T)被重载以表示具有相同值的对角矩阵；a＝[a 1,...,a K]：前K个最大的特征值的平均值a 1,...,a K的向量；
Figure PCTCN2019094919-appb-000012
前K个最大的特征值的平均值；Z∈ω(Λ)，ω(Λ)＝{Q TΛQ}；该方程可以通过迭代来求解。
具体求解过程可以包括：
步骤一、随机生成正交矩阵Q 0，Λ＝diag(λ)，
Figure PCTCN2019094919-appb-000013
且T＝[t ij]，其中，
Figure PCTCN2019094919-appb-000014
步骤二、在每次迭代中，T k由Z k-1生成。通过T k的特征分解生成Q k，如下：
Figure PCTCN2019094919-appb-000015
其中D＝diag(d)，d为T的前K个最大的特征值。
步骤三、根据以下计算方式更新Z k：
Figure PCTCN2019094919-appb-000016
T k由Z k-1生成，而Z k由T k生成，由此迭代求解得到正交矩阵Q。
最后的得到旋转后的投影矩阵P为：
P＝QW
通过正交旋转矩阵构建各向同性的低维空间。根据超球体的特性，将原始数据映射到一个各向同性的K维空间，然后在K维空间中应用球体散列，有效的降低了特征向量的维度，并由于各向同性的方差，使得超球体的构造精度更高，进而可以有效的提高二进制码计算相似图像的准确度和召回率。
在确定K个球体散列函数后，可以通过确定的K个球体散列函数，根据函数阈值，确定出待查询图像的二进制散列码中的K个比特位，以及图像库中存储图像的二进制散列码中的K个比特位。
针对线下预先根据图像库的存储图像确定K个球体散列函数的方式，可以在获得所述待查询图像之前，可以预先确定图像库中的存储图像的二进制码。在获得所述待查询图像之后，可根据预先确定的K个球体散列映射函数作为所述转换规则，确定所述待查询图像的K位二进制码。具体过程可以参考上述实施例，在此不再赘述。
在步骤103中，一种可能的实现方式，可以通过确定所述待查询图像的二进制码与所述各存储图像的二进制码的汉明距离，将汉明距离小于第一阈值的存储图像，确定为所述待查询图像的相似图像；
具体的，二进制码之间的距离度量可采用汉明距离、非对称距离等方式，在此不做限定。
在步骤104中，一种可能的实现方式，可以包括：
根据所述待查询图像的特征向量与所述相似图像的特征向量的欧式距离，确定所述待查询图像的特征向量与所述相似图像的相似度距离；
若确定在所述相似图像中，存在相似度距离满足预设条件的相似图像，则确定存在与所述待查询图像相似的存储图像，并将所述满足预设条件的存储图像作为所述待查询图像的查询结果；否则，所述查询结果为所述待查询图像不存在相似的存储图像。
根据本发明实施例中的图像检索方法，将图像库存储图像的特征向量、 待查询图像的特征向量作为做为输入数据，将特征向量映射为二进制码，大幅度减少图像的存储空间，并可以快速确定出所述待查询图像的相似图像。其中相似图像的张数以及映射后的比特数K是根据实际的图像库确定，在此不做限定。
举例来说，所述图像库以GIST1M数据集为例，该数据集包含100万个960维的GIST的特征向量。图2a、图2b、图2c所示，为将图像库分别映射为32bit、64bit和128bit时的召回率。其中iso-sph为本发明实施例中的算法，可以看到本发明实施例提供的图像检索方法，相比现有技术中的召回率更高，检索效果更好。并且将数据点映射为128bit时，查找待查询图像的前5000个最近邻召回率可以达到90％。
在GIST1M数据集使用本发明实施例提供的图像检索方法，检索一个待查询图像的前5000近邻所需时间如表1所示。可以看出，将数据映射为128bit时，在百万量级的图像库中检索一个待查询图像的前5000相似图像只需要9.70毫秒。
表1在GIST1M数据集上获取查询结果的响应时间(单位：毫秒)
  32bit 64bit 128bit
时间 3.82 4.96 9.70
将图像以二进制码表征时，存储空间将大幅度降低。假设每一张图片用一个4096维的单浮点数向量来，那么一个向量就需要16384个字节，也就是16千字节(KB)，如果用散列方法将每个高维向量映射成一个128bit(128bit＝16B)的散列码，节省约1000倍的存储空间。因此，本发明实施例中，在计算过程中无需将数据进行硬盘和内存的交换，可在内存中直接完成所有操作计算，提高计算效率。另外，使用本发明实施例中的图像检索方法，预先挑选出多个相似图像(例如，相似图像的个数小于10000),再对多个相似图像与待查询图片进行相似度测量，提高了后续待查询图像与图像库中的存储图像进行相似性计算的效率。
结合审核机构在审核用户资格等场景下，一种可能的实现方式，可以根据本发明实施例中的图像检索方法中确定出待查询图像的相似图像的步骤即可以提高审核效率。例如根据排序输出前10张待查询图像的相似图像，然后可以通过人工审核这10张相似图像，如有与待查询图像相同的图像则拒绝商户入网，如无相同的图像则同意商户入网。
为进一步提高审核效率，提高商户入网的体验，一种可能的实现方式，可以通过确定相似图像与待查询图像的相似度后，根据设定规则及图像查询结果，给出审核结果。例如，可以将待查询图像的特征向量与输出的相似图像的特征向量进行相似性距离测量，并设置审核阈值，当待查询图像的特征向量与相似图像的相似度距离小于审核阈值时，认为其相似度过高，则拒绝商户入网；否则，同意商户入网。通过上述方法，可以在确定商户提交资料有成功记录后，能立即给出审核通过的结果，进一步的提高了审核效率。
其中，本领域技术人员可以根据经验和实际情况设定预设阈值，具体不做限定。
本发明实施例还可以应用于人脸识别的应用场景中，以实现海量图库的人脸快速识别。对待查询图像进行特征提取后，通过本发明实施例中的方案，可以根据本发明实施例中的图像检索方法中确定出待查询图像的相似图像，再根据相似图像的特征向量与待查询图像的特征向量的相似度的比较，确定图像库中是否存在所述待查询图像，可以实现人脸识别的快速检索，到达毫秒级别的线上人脸识别速度，并能大幅度减少图像的存储空间。具体实施过程可以参考上述实施例，在此不再赘述。
需要说明的是，上述步骤编号仅为一种执行流程的示例性表示，本申请对各个步骤的先后顺序不做具体限定。
本发明提出了一种图像检索方法，解决人工审核千万量级图像资料效率低的问题；通过将图像的特征向量转换为二进制码，并通过比较二进制码间的相似度，确定有效近邻图像，大幅度减少相似度测量的有效近邻图像数量，再根据邻近图像的特征向量与待查询图像的特征向量进行相似度计算，获得 高精度的检索结果，有效提高了图像的检索速度。并且，由于将图像的特征向量转还为二进制码与图像库中的图像的二进制码进行比对，可以极大的降低图像特征向量在检索过程中占用的存储空间，实现单机的图像检索，图像数据无需进行内存与硬盘之间的数据交换，即在单机内存中能进行检索的所有操作计算。在检索过程中，不需要对图库进行标签，可以适用于无标签的图库，极大的降低人工标注成本，减少模型训练成本。
基于同样的发明构思，图3示例性示出了本发明实施例提供的一种图像检索装置的结构示意图，如图3所示，该装置包括获取模块301、处理模块302。
本发明实施例提供一种图像检索装置，所述装置包括：
获取模块301，用于确定待查询图像的特征向量；
处理模块302，用于将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的；将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。
一种可能的实现方式，所述处理模块302，还用于：
根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间的各投影方向的方差相同且最大；所述各参照图像包括所述各存储图像和所述待查询图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种可能的实现方式，所述转换规则通过以下方式确定：根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大且相同；所述各参照图像为所述各存储图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
一种可能的实现方式，所述处理模块302，具体用于：
确定将所述各参照图像的特征向量投影至K维空间的投影矩阵，以使所述各参照图像的特征向量在所述K维空间的各投影方向上的投影距离的方差最大；根据正交矩阵旋转所述投影矩阵，以使所述各参照图像的特征向量在旋转后的K维空间的各投影方向上的投影距离的方差相同；根据旋转后的投影矩阵，确定所述K维空间。
一种可能的实现方式，所述处理模块302，具体用于：
通过确定所述待查询图像的二进制码与所述各存储图像的二进制码的汉明距离，将汉明距离小于第一阈值的存储图像，确定为所述待查询图像的相似图像；根据所述待查询图像的特征向量与所述相似图像的特征向量的欧式距离，确定所述待查询图像的特征向量与所述相似图像的相似度距离；若确定在所述相似图像中，存在相似度距离满足预设条件的相似图像，则确定存在与所述待查询图像相似的存储图像，并将所述满足预设条件的存储图像作为所述待查询图像的查询结果；否则，所述查询结果为所述待查询图像不存在相似的存储图像。
基于相同的技术构思，本申请实施例提供了一种图像检索的设备，如图4所示，包括至少一个存储器401，以及与至少一个存储器连接的处理器402，本申请实施例中不限定处理器402与存储器401之间的具体连接介质，图4中处理器402和存储器401之间通过总线连接为例。总线可以分为地址总线、数据总线、控制总线等。
进一步地，本发明实施例中所述的计算设备还可以包括输入装置403以及输出装置404等。输入装置403可以包括键盘、鼠标、触摸屏等；输出装置404可以包括显示设备，如液晶显示器(Liquid Crystal Display，LCD)、阴极射线管(Cathode Ray Tube，CRT)，触摸屏等。
在本申请实施例中，存储器401存储有可被至少一个处理器402执行的指令，至少一个处理器402通过执行存储器401存储的指令，可以执行前述的图像检索的方法中所包括的步骤。
其中，处理器402是图像检索的设备的控制中心，可以利用各种接口和 线路连接图像检索的设备的各个部分，通过运行或执行存储在存储器401内的指令以及调用存储在存储器401内的数据，从而实现图像检索。可选的，处理器402可包括一个或多个处理单元，处理器402可集成应用处理器和调制解调处理器，其中，应用处理器主要处理操作系统、用户界面和应用程序等，调制解调处理器主要处理无线通信。可以理解的是，上述调制解调处理器也可以不集成到处理器402中。在一些实施例中，处理器402和存储器401可以在同一芯片上实现，在一些实施例中，它们也可以在独立的芯片上分别实现。
处理器402可以是通用处理器，例如中央处理器(CPU)、数字信号处理器、专用集成电路(AppKication Specific Integrated Circuit，ASIC)、现场可编程门阵列或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本申请实施例中公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者任何常规的处理器等。结合本申请实施例所公开的方法的步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。
存储器401作为一种非易失性计算机可读存储介质，可用于存储非易失性软件程序、非易失性计算机可执行程序以及模块。存储器401可以包括至少一种类型的存储介质，例如可以包括闪存、硬盘、多媒体卡、卡型存储器、随机访问存储器(Random Access Memory，RAM)、静态随机访问存储器(Static Random Access Memory，SRAM)、可编程只读存储器(ProgrammabKe Read OnKy Memory，PROM)、只读存储器(Read OnKy Memory，ROM)、带电可擦除可编程只读存储器(EKectricaKKy ErasabKe ProgrammabKe Read-OnKy Memory，EEPROM)、磁性存储器、磁盘、光盘等等。存储器401是能够用于携带或存储具有指令或数据结构形式的期望的程序代码并能够由计算机存取的任何其他介质，但不限于此。本申请实施例中的存储器401还可以是电路或者其它任意能够实现存储功能的装置，用于存储程序指令和/或数据。
基于同一发明构思，本申请实施例还提供了一种计算机可读介质，其存 储有可由图像检索的设备执行的计算机程序，当所述程序在图像检索的设备上运行时，使得所述图像检索的设备执行图像检索的方法的步骤。
本领域内的技术人员应明白，本发明的实施例可提供为方法、或计算机程序产品。因此，本发明可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本发明可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本领域内的技术人员应明白，本发明的实施例可提供为方法、系统、或计算机程序产品。因此，本发明可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本发明可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图 一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本发明的优选实施例，但本领域内的技术人员一旦得知了基本创造性概念，则可对这些实施例作出另外的变更和修改。所以，所附权利要求意欲解释为包括优选实施例以及落入本发明范围的所有变更和修改。
显然，本领域的技术人员可以对本发明进行各种改动和变型而不脱离本发明的精神和范围。这样，倘若本发明的这些修改和变型属于本发明权利要求及其等同技术的范围之内，则本发明也意图包含这些改动和变型在内。

Claims (12)

  1. 一种图像检索方法，其特征在于，所述方法包括：
    确定待查询图像的特征向量；
    将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的；
    将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；
    针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。
  2. 根据权利要求1所述的方法，其特征在于，所述将所述待查询图像的特征向量依据转换规则转换为二进制码之前，还包括：
    根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间的各投影方向的方差相同且最大；所述各参照图像包括所述各存储图像和所述待查询图像；K为正整数；
    在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
  3. 根据权利要求1所述的方法，其特征在于，所述转换规则通过以下方式确定：
    根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大且相同；所述各参照图像为所述各存储图像；
    在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
  4. 根据权利要求2或3所述的方法，其特征在于，所述根据各参照图像确定K维空间，包括：
    确定将所述各参照图像的特征向量投影至K维空间的投影矩阵，以使所述各参照图像的特征向量在所述K维空间的各投影方向上的投影距离的方差最 大；
    根据正交矩阵旋转所述投影矩阵，以使所述各参照图像的特征向量在旋转后的K维空间的各投影方向上的投影距离的方差相同；
    根据旋转后的投影矩阵，确定所述K维空间。
  5. 根据权利要求1所述的方法，其特征在于，所述将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像，包括：
    通过确定所述待查询图像的二进制码与所述各存储图像的二进制码的汉明距离，将汉明距离小于第一阈值的存储图像，确定为所述待查询图像的相似图像；
    所述针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果，包括：
    根据所述待查询图像的特征向量与所述相似图像的特征向量的欧式距离，确定所述待查询图像的特征向量与所述相似图像的相似度距离；
    若确定在所述相似图像中，存在相似度距离满足预设条件的相似图像，则确定存在与所述待查询图像相似的存储图像，并将所述满足预设条件的存储图像作为所述待查询图像的查询结果；否则，所述查询结果为所述待查询图像不存在相似的存储图像。
  6. 一种图像检索装置，其特征在于，所述装置包括：
    获取模块，用于确定待查询图像的特征向量；
    处理模块，用于将所述待查询图像的特征向量依据转换规则转换为二进制码，所述转换规则是至少根据图像库中各存储图像确定的；将所述待查询图像的二进制码与所述图像库中各存储图像的二进制码进行比较，确定所述待查询图像的相似图像；针对每张相似图像，根据所述相似图像的特征向量与所述待查询图像的特征向量间的相似度，确定所述待查询图像的查询结果。
  7. 根据权利要求6所述的装置，其特征在于，所述处理模块，还用于：
    根据各参照图像确定K维空间，在所述各参照图像的特征向量在K维空间 的各投影方向的方差相同且最大；所述各参照图像包括所述各存储图像和所述待查询图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
  8. 根据权利要求7所述的装置，其特征在于，所述转换规则通过以下方式确定：根据各参照图像确定K维空间，所述各参照图像的特征向量在所述K维空间的投影距离的方差最大且相同；所述各参照图像为所述各存储图像；在所述K维空间中，求解所述各参照图像的特征向量的K个球体散列映射函数；所述K个球体散列映射函数作为所述转换规则。
  9. 根据权利要求7或8所述的装置，其特征在于，所述处理模块，具体用于：
    确定将所述各参照图像的特征向量投影至K维空间的投影矩阵，以使所述各参照图像的特征向量在所述K维空间的各投影方向上的投影距离的方差最大；根据正交矩阵旋转所述投影矩阵，以使所述各参照图像的特征向量在旋转后的K维空间的各投影方向上的投影距离的方差相同；根据旋转后的投影矩阵，确定所述K维空间。
  10. 根据权利要求6所述的装置，其特征在于，所述处理模块，具体用于：
    通过确定所述待查询图像的二进制码与所述各存储图像的二进制码的汉明距离，将汉明距离小于第一阈值的存储图像，确定为所述待查询图像的相似图像；根据所述待查询图像的特征向量与所述相似图像的特征向量的欧式距离，确定所述待查询图像的特征向量与所述相似图像的相似度距离；若确定在所述相似图像中，存在相似度距离满足预设条件的相似图像，则确定存在与所述待查询图像相似的存储图像，并将所述满足预设条件的存储图像作为所述待查询图像的查询结果；否则，所述查询结果为所述待查询图像不存在相似的存储图像。
  11. 一种图像检索的设备，其特征在于，包括至少一个处理器、以及至少一个存储器，其中，所述存储单元存储有计算机程序，当所述程序被所述处理器执行时，使得所述处理器执行权利要求1～5任一权利要求所述方法的 步骤。
  12. 一种计算机可读介质，其特征在于，其存储有可由图像检索的设备执行的计算机程序，当所述程序在图像检索的设备上运行时，使得所述图像检索的设备执行权利要求1～5任一所述方法的步骤。
PCT/CN2019/094919 2018-11-15 2019-07-05 一种图像检索方法及装置 Ceased WO2020098296A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811358729.0A CN109522435B (zh) 2018-11-15 2018-11-15 一种图像检索方法及装置
CN201811358729.0 2018-11-15

Publications (1)

Publication Number Publication Date
WO2020098296A1 true WO2020098296A1 (zh) 2020-05-22

Family

ID=65777930

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/094919 Ceased WO2020098296A1 (zh) 2018-11-15 2019-07-05 一种图像检索方法及装置

Country Status (2)

Country Link
CN (1) CN109522435B (zh)
WO (1) WO2020098296A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113609165A (zh) * 2021-08-13 2021-11-05 广联达科技股份有限公司 数据请求方法、数据处理方法和数据管理系统
CN114973295A (zh) * 2021-02-25 2022-08-30 深圳富联富桂精密工业有限公司 相似机构件查找方法、装置及计算机可读存储介质
CN115481276A (zh) * 2022-09-21 2022-12-16 北京百度网讯科技有限公司 基于人工智能的图像检索方法、装置、设备及存储介质
CN115878824A (zh) * 2020-07-07 2023-03-31 华为技术有限公司 图像检索系统、方法和装置
WO2024119937A1 (zh) * 2022-12-08 2024-06-13 华为技术有限公司 一种图像检索的方法、系统以及装置

Families Citing this family (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109522435B (zh) * 2018-11-15 2022-05-20 中国银联股份有限公司 一种图像检索方法及装置
CN111930984A (zh) * 2019-04-24 2020-11-13 北京京东振世信息技术有限公司 图像检索方法、装置、服务器、客户端及介质
CN110765292A (zh) * 2019-10-24 2020-02-07 重庆紫光华山智安科技有限公司 图像检索方法、训练方法及相关装置
CN110826627A (zh) * 2019-11-06 2020-02-21 广东三维家信息科技有限公司 图像相似度测量方法、装置及电子设备
CN113094530B (zh) * 2019-12-23 2024-03-29 深圳云天励飞技术有限公司 一种图像数据检索方法、装置、电子设备及存储介质
CN113094539A (zh) * 2019-12-23 2021-07-09 深圳云天励飞技术有限公司 一种图像数据的搜索方法、装置、电子设备及存储介质
CN111125417B (zh) * 2019-12-30 2023-03-31 深圳云天励飞技术有限公司 一种数据搜索方法、装置、电子设备及存储介质
CN110929706B (zh) * 2020-02-19 2020-06-02 北京海天瑞声科技股份有限公司 视频抽选方法、装置及存储介质
CN113297409A (zh) * 2021-06-30 2021-08-24 杭州海康威视数字技术股份有限公司 一种图像搜索方法、装置、电子设备及存储介质
CN115062180B (zh) * 2022-08-15 2023-03-28 阿里巴巴（中国）有限公司 对象查询的方法、电子设备及存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101710334A (zh) * 2009-12-04 2010-05-19 大连理工大学 基于图像哈希的大规模图像库检索方法
CN103226585A (zh) * 2013-04-10 2013-07-31 大连理工大学 面向图像检索的自适应哈希重排方法
CN104462503A (zh) * 2014-12-19 2015-03-25 中国科学院深圳先进技术研究院 确定数据点的相似度的方法
US20160267637A1 (en) * 2015-03-12 2016-09-15 Yahoo! Inc. System and method for improved server performance for a deep feature based coarse-to-fine fast search
CN106777349A (zh) * 2017-01-16 2017-05-31 广东工业大学 基于深度学习的人脸检索系统及方法
CN106909924A (zh) * 2017-02-18 2017-06-30 北京工业大学 一种基于深度显著性的遥感影像快速检索方法
CN109522435A (zh) * 2018-11-15 2019-03-26 中国银联股份有限公司 一种图像检索方法及装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101710334A (zh) * 2009-12-04 2010-05-19 大连理工大学 基于图像哈希的大规模图像库检索方法
CN103226585A (zh) * 2013-04-10 2013-07-31 大连理工大学 面向图像检索的自适应哈希重排方法
CN104462503A (zh) * 2014-12-19 2015-03-25 中国科学院深圳先进技术研究院 确定数据点的相似度的方法
US20160267637A1 (en) * 2015-03-12 2016-09-15 Yahoo! Inc. System and method for improved server performance for a deep feature based coarse-to-fine fast search
CN106777349A (zh) * 2017-01-16 2017-05-31 广东工业大学 基于深度学习的人脸检索系统及方法
CN106909924A (zh) * 2017-02-18 2017-06-30 北京工业大学 一种基于深度显著性的遥感影像快速检索方法
CN109522435A (zh) * 2018-11-15 2019-03-26 中国银联股份有限公司 一种图像检索方法及装置

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115878824A (zh) * 2020-07-07 2023-03-31 华为技术有限公司 图像检索系统、方法和装置
CN115878824B (zh) * 2020-07-07 2023-10-20 华为技术有限公司 图像检索系统、方法和装置
CN114973295A (zh) * 2021-02-25 2022-08-30 深圳富联富桂精密工业有限公司 相似机构件查找方法、装置及计算机可读存储介质
CN113609165A (zh) * 2021-08-13 2021-11-05 广联达科技股份有限公司 数据请求方法、数据处理方法和数据管理系统
CN115481276A (zh) * 2022-09-21 2022-12-16 北京百度网讯科技有限公司 基于人工智能的图像检索方法、装置、设备及存储介质
WO2024119937A1 (zh) * 2022-12-08 2024-06-13 华为技术有限公司 一种图像检索的方法、系统以及装置
CN119293278A (zh) * 2022-12-08 2025-01-10 华为技术有限公司 一种图像检索的方法、系统以及装置

Also Published As

Publication number Publication date
CN109522435A (zh) 2019-03-26
CN109522435B (zh) 2022-05-20

Similar Documents

Publication Publication Date Title
CN109522435B (zh) 一种图像检索方法及装置
Zhang et al. Global context aware convolutions for 3d point cloud understanding
US11593673B2 (en) Systems and methods for identifying influential training data points
US20160203191A1 (en) Recommendation system with metric transformation
Ni et al. An improved adaptive ORB-SLAM method for monocular vision robot under dynamic environments
CN111428457A (zh) 数据表的自动格式化
CN104484671B (zh) 应用于移动平台的目标检索系统
CN106447592A (zh) 每个特征描述符的在线个性化服务
CN114419406A (zh) 图像变化检测方法、训练方法、装置和计算机设备
CN110956131A (zh) 单目标追踪方法、装置及系统
CN114218397A (zh) 交易关系图谱处理方法、装置、计算机设备及存储介质
CN118626301A (zh) 进程执行异常问题定位方法、装置、设备、介质和产品
CN108229572B (zh) 一种参数寻优方法及计算设备
CN116894964B (zh) 后期融合物体图像聚类方法、装置和计算机设备
CN111695917B (zh) 商品推荐方法、系统、电子设备和存储介质
CN116894218B (zh) 基于显著性正则化图神经网络的图分类方法及相关装置
WO2025246128A1 (zh) 小样本人脸美丽度评价方法、设备及介质
Zhou et al. Inexact Riemannian Gradient Descent Method for Nonconvex Optimization with Strong Convergence: J. Zhou et al.
CN114743150B (zh) 目标跟踪方法、装置、电子设备及存储介质
CN115758180A (zh) 基于拓补结构的用户评分方法、系统、设备及存储介质
CN107885757B (zh) 图像检索的方法及装置
CN119722756B (zh) 三维点云配准方法、装置及电子设备
CN117892166B (zh) 稀疏数据识别方法、智能计算子系统以及智能计算平台
Zhu et al. Branched convolutional neural networks for face alignment
US20240386317A1 (en) Systems and methods for configuring and training machine learning models based on input value characteristics

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19885526

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19885526

Country of ref document: EP

Kind code of ref document: A1