(19)
(11)EP 3 683 724 A1

(12)EUROPEAN PATENT APPLICATION
published in accordance with Art. 153(4) EPC

(43)Date of publication:
22.07.2020 Bulletin 2020/30

(21)Application number: 18857317.4

(22)Date of filing:  07.09.2018
(51)International Patent Classification (IPC): 
G06K 9/62(2006.01)
(86)International application number:
PCT/CN2018/104565
(87)International publication number:
WO 2019/052403 (21.03.2019 Gazette  2019/12)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
KH MA MD TN

(30)Priority: 12.09.2017 CN 201710817344

(71)Applicant: Tencent Technology (Shenzhen) Company Limited
Shenzhen, Guangdong 518057 (CN)

(72)Inventors:
  • MA, Lin
    Shenzhen City Guangdong 518057 (CN)
  • JIANG, Wenhao
    Shenzhen City Guangdong 518057 (CN)
  • LIU, Wei
    Shenzhen City Guangdong 518057 (CN)

(74)Representative: AWA Sweden AB 
P.O. Box 45086
104 30 Stockholm
104 30 Stockholm (SE)

  


(54)TRAINING METHOD FOR IMAGE-TEXT MATCHING MODEL, BIDIRECTIONAL SEARCH METHOD, AND RELATED APPARATUS


(57) This application relates to the field of artificial intelligence technologies, and in particular, to a training method for an image-text matching model, a bidirectional search method, and a related apparatus. The training method includes: extracting global representations and local representations of an image sample and a text sample; and training a matching model, to determine model parameters of the matching model, the matching model being used to determine, according to a global representation and a local representation of an image and a global representation and a local representation of a text, a matching degree between the image and the text. The matching degree in this application is obtained based on all of the detailed features and the global feature of the image, and the obtained matching degree is more accurate and comprehensive.




Description

RELATED APPLICATION



[0001] This application claims priority to Chinese Patent Application No. 201710817344.5, entitled "TRAINING METHOD FOR IMAGE-TEXT MATCHING MODEL, BIDIRECTIONAL SEARCH METHOD, AND RELATED APPARATUS" filed with the China National Intellectual Property Administration on September 12, 2017, which is incorporated by reference in its entirety.

FIELD OF THE TECHNOLOGY



[0002] This application relates to the field of artificial intelligence technologies, and in particular, to a training method for an image-text matching model, a bidirectional search method, and a related apparatus.

BACKGROUND



[0003] Understanding of images and texts has always been one of the most important research directions in artificial intelligence. One important research is exploring relationships between images and texts. For example, news text content and a news image in web page news may express the same theme. That is, there is no absolute independence between images and texts, and there are matching relationships between the images and the texts. Therefore, how to find a text matching a given image or find an image matching a given text becomes a topic of interest in the industry.

SUMMARY



[0004] Embodiments of this application provide a training method for an image-text matching model and a related apparatus, so that a matching degree between an image and a text can be comprehensively measured, and the obtained matching degree is more accurate and comprehensive.

[0005] The embodiments of this application provide a bidirectional search method and a related apparatus, so that when an image-text bidirectional search is performed, a more accurate search result can be obtained.

[0006] The embodiments of this application provide a training method for an image-text matching model, the method being carried out by a computer and including:

extracting a global feature and a local feature of an image sample;

extracting a global feature and a local feature of a text sample; and

training a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an image and a global feature and a local feature of a text, a matching degree between the image and the text.



[0007] The embodiments of this application provide an image-text bidirectional search method, the method being carried out by a computer and including:

receiving a reference sample, the reference sample being a text or an image;

extracting a global feature and a local feature of the reference sample;

inputting the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample and a global feature and a local feature of the corresponding material; and

selecting a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.



[0008] The embodiments of this application provide a training apparatus for an image-text matching model, including:

an image feature extraction module, configured to extract a global feature and a local feature of an image sample;

a text feature extraction module, configured to extract a global feature and a local feature of a text sample; and

a training module, configured to train a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an image and a global feature and a local feature of a text, a matching degree between the image and the text.



[0009] The embodiments of this application provide an image-text bidirectional search apparatus, including:

a reference sample receiving module, configured to receive a reference sample, the reference sample being a text or an image;

a reference sample feature extraction module, configured to extract a global feature and a local feature of the reference sample;

a search module, configured to input the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample and a global feature and a local feature of the corresponding material; and

a selection module, configured to select a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.



[0010] Embodiments of this application provide a computing device, including a memory and a processor, the memory being configured to store program instructions, and the processor being configured to invoke the program instructions stored in the memory, to perform the training method for an image-text matching model according to any one of the embodiments of this application according to the program instructions.

[0011] The embodiments of this application provide a computer storage medium, storing computer-executable instructions, the computer-executable instructions being used to enable the computer to perform the training method for an image-text matching model according to any one the embodiments of this application according to the computer-executable instructions.

[0012] The embodiments of this application provide a computing device, including a memory and a processor, the memory being configured to store program instructions, and the processor being configured to invoke the program instructions stored in the memory, to perform the image-text bidirectional search method according to the embodiments of this application according to the program instructions.

[0013] The embodiments of this application provide a computer storage medium, storing computer-executable instructions, the computer-executable instructions being used to enable the computer to perform the image-text bidirectional search method according to the embodiments of this application according to the computer-executable instructions.

BRIEF DESCRIPTION OF THE DRAWINGS



[0014] 

FIG. 1 is a schematic diagram of an application scenario according to some embodiments of this application.

FIG. 2A is a schematic flowchart of a training method for an image-text matching model according to some embodiments of this application.

FIG. 2B is a schematic flowchart of a training method for an image-text matching model according to some embodiments of this application.

FIG. 3 is a block diagram of a matching model according to some embodiments of this application.

FIG. 4 is a block diagram of a model for calculating a similarity of an image feature according to some embodiments of this application.

FIG. 5 is a schematic diagram of a category of image information according to some embodiments of this application.

FIG. 6 is a block diagram of extracting a global representation of a text according to some embodiments of this application.

FIG. 7 is a schematic flowchart of a training method for an image-text matching model according to some embodiments of this application.

FIG. 8 is a schematic flowchart of an image-text bidirectional search method according to some embodiments of this application.

FIG. 9 is a schematic structural diagram of a training apparatus for an image-text matching model according to some embodiments of this application.

FIG. 10 is a schematic structural diagram of an image-text bidirectional search apparatus according to some embodiments of this application.

FIG. 11 is a schematic structural diagram of a computer device according to some embodiments of this application.


DESCRIPTION OF EMBODIMENTS



[0015] For ease of understanding the technical solution provided by the embodiments of this application, the embodiments of this application are described below in detail with reference to the accompanying drawings of the specification.

[0016] In some embodiments, matching between an image and a text is implemented through two methods.

Method 1: Extract respective global representations of an image and a text, map the global representations to the same semantic space, and establish a matching relationship between the global representations of the image and the text in the semantic space, so as to establish a matching relationship between the image and the text. The global representations reflect global characteristics of the image, and are used to describe overall features of the image, and the global representations are, for example, a color feature, a shape feature, a texture feature, and the like.

Method 2: Obtain a local representation of local information of an image through a convolutional neural network, and obtain a local representation of a text through syntax tree information, so as to learn a matching relationship between the local representations of the image and the text. The local representation reflects local information of the image, and is used to describe local detailed features of the image. Compared with the global representation, the local representation is characterized by a large quantity in an image, a small relevance between features, and detection and matching of other features would not be affected by disappearance of some local features in an obscured situation.



[0017] However, in the foregoing two methods, matching between an image and a text is performed by using information in a single aspect, that is, either independently using a global representation or independently using a local representation. Whether the local representation or the global representation is used to perform matching between an image and a text, a matching degree between the image and the text can only be reflected on one level. For an image with rich contents, by using the global representation, only color features, shape features, and texture features can be extracted, but the detailed features of the local representation cannot be represented. Consequently, the features that can be reflected are limited. By using the local representation, a matching relationship between an image and a text can only be considered locally, and a feature of the global representation cannot be reflected. Therefore, neither of the foregoing methods can comprehensively measure a matching degree between an image and a text.

[0018] To comprehensively match an image and a text and obtain a more accurate matching degree, in an embodiment of this application, an image-text matching model is trained according to the following method, including:

extracting a global feature and a local feature of an image sample;

extracting a global feature and a local feature of a text sample; and

training a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an image and a global feature and a local feature of a text, a matching degree between the image and the text.



[0019] In this way, the matching degree is obtained based on the respective global representations and local representations of the image and the text. Since the detailed features and the global feature of the image are all considered, the obtained matching degree is more accurate and comprehensive. Therefore, when an image-text bidirectional search is performed based on the matching model provided by this embodiment of this application, a search result is also more accurate.

[0020] For example, referring to FIG. 1, FIG. 1 is a schematic diagram of an application scenario according to an embodiment of this application. The scenario, for example, may include a terminal device 11 and a server 12. Various clients, such as a blog client for social communication, a WeChat client, and a news client for learning current events, may be installed in the terminal device 11. After a client of the terminal device 11 and the server 12 establish a communication connection, the client of the terminal device 11 may extract global features and local features of the image and/or the text and transmit them to the server 12. Alternatively, the client of the terminal device 11 transmits the image and/or text to the server 12, and the server 12 extracts the global features and local features of the image and/or the text. Then, the server 12 searches for a text matching the image according to the global feature and the local feature of the image and based on the matching model, or the server 12 searches for an image matching the text according to the global feature and the local feature of the text and based on the matching model. Subsequently, the server 12 returns a search result to the terminal device 11.

[0021] The terminal device 11 and the server 12 may be communicatively connected through a communications network. The communication network may be a local area network, a wide area network, or the like. The terminal device 11 may be a mobile phone, a tablet computer, a notebook computer, a personal computer, or the like. The server 12 may be any server device capable of supporting training of an image-text matching model and a bidirectional search.

[0022] The solution provided by this embodiment of this application is applicable to any scenario for searching for an image according to a text and/or searching for a text according to an image. For example, the solution may be applicable to a scenario for searching for an image and/or a text in cloud album management (for example, searching for a matching image according to a given text), WeChat Moments, QQ Space, QQ Mass Innovation Space, and chat environments of WeChat and QQ.

[0023] Referring to FIG. 2A, FIG. 2A is a schematic flowchart of a training method for an image-text matching model according to an embodiment of this application. The method is carried out by a computer, and the computer may be a server 12. The method includes following steps.

S21: Extract a global feature and a local feature of an image sample.

S22: Extract a global feature and a local feature of a text sample.

S23: Train a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an image and a global feature and a local feature of a text, a matching degree between the image and the text.



[0024] When this technical solution provided by this application is used, the matching model is trained according to the global features and the local features of the image sample and the text sample. Since the local features and the global features of the image and the text are all considered, a matching degree determined according to the trained matching model is more accurate and comprehensive. When an image-text bidirectional search is performed based on the matching model provided by this embodiment of this application, a search result is more accurate.

[0025] Referring to FIG. 2B, FIG. 2B is a flowchart of a training method for an image-text matching model according to an embodiment of this application. The method may be carried out by the server 12 shown in FIG. 1, and includes the following steps.

[0026] Step 201: Extract a global feature and a local feature of an image sample.

[0027] In some embodiments, the global feature is a vector, which is used to represent text information of a corresponding image. The local feature of the image sample is also a vector, and information of each dimension of the vector represents that a local area of the image includes information of a corresponding word category, for example, a probability that a local area of the image includes information of a corresponding word category.

[0028] Step 202: Extract a global feature and a local feature of a text sample.

[0029] In some embodiments, the global feature of the text sample is used to represent overall semantic information of the text sample. The local feature of the text sample is used to represent synthesis of local semantic information of the text. The local feature of the text sample focuses on local information of the text, such as a word or a phrase.

[0030] The execution sequence of step 201 and step 202 is not limited to the description above.

[0031] Step 203: Train a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an inputted image and a global feature and a local feature of an inputted text, a matching degree between the image and the text.

[0032] Next, after the matching model is completely trained, the matching degree between the image and the text may be determined according to the trained matching model, including the following steps.

[0033] Step 204: Map the respective global features of the image and the text to a specified semantic space through the matching model, and calculate a similarity between the global features of the image and the text; map the respective local features of the image and the text to the specified semantic space, and calculate a similarity between the local features of the image and the text; and determine the matching degree between the image and the text using a weighted summation manner according to a preset weight of the similarity between the global features and a preset weight of the similarity between the local features.

[0034] As shown in FIG. 3, FIG. 3 is a block diagram of an image-text matching model according to an embodiment of this application. In FIG. 3, for an inputted image 310, a local image convolutional neural network (CNN) 311 is used to extract a local representation (which is also referred to as a local feature) of the image 310 (as shown by 312 in FIG. 3), and a global image CNN 313 is used to extract a global representation (which is also referred to as a global feature) of the image 310 (as shown by 314 in FIG. 3). For an inputted text 320, a local text encoder 321 is used to extract a local representation (which is also referred to as a local feature) of the text 320 (as shown by 322 in FIG. 3), and a global text encoder 323 is used to extract a global representation (which is also referred to as a global feature) of the text 320 (as shown by 324 in FIG. 3). After the local representations and the global representations of the image 310 and the text 320 are extracted, the respective local representations of the image 310 and the text 320 are mapped to a specified semantic space through local depth matching, and a similarity between local features of a heterogeneous sample pair (which is also referred to as an image-text pair) constituted by the image 310 and the text 320 is determined in the specified semantic space. Similarly, the respective global features of the image 310 and the text 320 are mapped to the specified semantic space through global depth matching, and a similarity between global features of the heterogeneous sample pair constituted by the image 310 and the text 320 is determined in the specified semantic space. Subsequently, through model integration, a matching degree of the heterogeneous sample pair is determined using a weighted summation manner according to a preset weight of the similarity between the global features and a preset weight of the similarity between the local features.

[0035] For ease of understanding, an embodiment of this application provides a training method for an image-text matching model, which is described by using the following contents.

[0036] 1) With regard to mapping to a specified semantic space.

[0037] In some embodiments, to obtain an accurate similarity between the global representations of the image and the text and an accurate similarity between the local representations of the image and the text, the mapping the respective global representations of the image sample and the text sample are mapped to the specified semantic space includes:

respectively mapping the respective global features of the image sample and the text sample to the specified semantic space through at least two fully connected layers; and

the mapping the respective local representations of the image sample and the text sample to the specified semantic space includes:
respectively mapping the respective local features of the image sample and the text sample to the specified semantic space through the at least two fully connected layers.



[0038] The fully connected layer is a layer in a basic structure of the convolutional neural network. Each node of the fully connected layer is connected to all nodes of a previous layer to integrate the extracted features. The fully connected layer is configured to map the global features and the local features of the image and the text to the specified semantic space. A full connection operation refers to an operation of mapping to a specified semantic space through a fully connected layer. Parameters of the fully connected layer are obtained through training according to a preset target function. The preset target function is used to determine that a similarity of a semantically associated heterogeneous sample pair (image-text pair) is higher than a similarity of a non-semantically associated heterogeneous sample pair. Alternatively, the preset target function is used to determine that a similarity of a semantically associated heterogeneous sample pair is higher than a similarity of a non-semantically associated heterogeneous sample pair, and that a similarity between text samples associated with the same image sample is higher than a similarity between text samples associated with different image samples.

[0039] In this embodiment of this application, the respective global representations and local representations of the text and the image are accurately analyzed through a plurality of fully connected layers, so as to map the global representations and the local representations of the text and the image to the same semantic space. In addition, the parameters of the fully connected layers are determined according to the preset target function. A similarity of a heterogeneous sample pair can be accurately determined through the preset target function according to an actual situation, wherein the similarity is a similarity of global representations or a similarity of local representations.

[0040] In addition, multi-layer nonlinear transformation may be performed on descriptions of the text and the image through the plurality of fully connected layers, so as to explore a direct matching relationship between the image and the text.

[0041] If the preset target function is used to realize that a similarity of a semantically associated heterogeneous sample pair is higher than a similarity of a non-semantically associated heterogeneous sample pair, the preset target function may be shown in formula (1):

where

represents the preset target function;

and

represent the text sample;

and

represent the image sample; d() represents a similarity between global representations of the text sample and the image sample when a similarity between global representations is determined; d() represents a similarity between local representations of the text sample and the image sample when a similarity between local representations is determined; λ1 and λ2λ2 both represent preset coefficients;

represents the similarity of the semantically associated heterogeneous sample pair;

and

both represent the similarity of the non-semantically associated heterogeneous sample pair; and µ1 and µ2 both represent preset thresholds.

[0042] It is considered that one image is generally semantically associated with a plurality of samples. For example, in a landscape image where a bird flies in the sky and a fish swims in the water, corresponding semantic information may include that the bird flies in the sky and the fish swims in the water. Therefore, in order to improve accuracy of a matching model, it is necessary to make a similarity between texts semantically associated with the same image higher than a similarity between the texts semantically associated with other different images. On the basis of satisfying the condition as described in formula (1), the preset target function also needs to satisfy that a similarity between text samples associated with the same image sample is higher than a similarity between text samples associated with different image samples. To reach the objective, a formula for making the similarity between text samples associated with the same image sample higher than the similarity between text samples associated with different image samples is shown as formula (2):

where L(Si, Sl,Sj) represents the preset target function that requires the similarity between text samples associated with the same image sample to be higher than the similarity between text samples associated with different image samples; Si,Sl Si,Sl represent the text samples semantically associated with the same image sample; Si , Sj Si,Sj represent the text samples associated with different image samples; d()d() represents a similarity between global representations of the text samples when a similarity between global representations is determined; d()d() represents a similarity between local representations of the text samples when a similarity between local representations is determined; and u3 u3 represents a preset threshold.

[0043] In some embodiments, the formula (1) and the formula (2) may be integrated, to obtain a preset target function as described in the formula (3) or the formula (4), and the obtained preset target function can satisfy the requirements as described in the formula (1) and the formula (2):

and



[0044] In an embodiment, a specific mathematical expression manner may be designed according to the requirements of the preset target function. This is not limited in this embodiment of this application.

[0045] According to an effect of the preset target function, a depth matching model as shown in FIG. 4 may be trained. The depth matching model is used to determine a matching degree between an image and a text. In FIG. 4, a feature vector (image embedding) of an image (the feature vector is a feature vector of a global feature or a local feature) is inputted on a left side, and a feature vector (sentence embedding) of a text (the feature vector is a feature vector of a global feature or a feature vector of a local feature) is inputted on a right side. After the feature vectors of the image and the text are mapped to the same specified semantic space, a distance between the feature vectors is calculated. For example, in FIG. 4,







and Sm represent the feature vectors of the text, and I1 and In represent the feature vectors of the image. In the specified semantic space, a distance between the image and the text is calculated. If the feature vectors used are feature vectors for representing global features, a calculation result is a similarity between the global features. If the feature vectors used are feature vectors for representing local features, a calculation result is a similarity between the local features.

[0046] 2) With regard to extraction of a global representation (global feature) and a local representation (local feature) of an image sample

[0047] The extraction of a global feature of an image is not limited in this embodiment of this application. For example, a global representation of an image is represented by using a feature of a fully connected layer of a convolutional neural network of the image, for example, the foregoing CNN. The CNN herein may be selected from, but is not limited to, a Very Deep Convolutional Networks for Large-Scale Image Recognition of Visual Geometry Group (VGG), a Residual Network (ResNet), Inception, Inception V3, and Inception V4. Certainly, network models, such as Fully Convolutional Networks (FCNs) for Semantic Segmentation, may alternatively be used to extract a global representation of an image.

[0048] Herein, the extraction of a local representation of the image sample provided in this embodiment of this application is described, and may be executed as follows:

dividing the image sample into a specified quantity of image blocks, and for each image block, calculating a probability that the image block includes a specified category of image information; and

selecting maximum probabilities of respective specified categories of image information in the specified quantity of image blocks, the maximum probabilities of the respective specified categories of image information constituting the local feature of the image sample.



[0049] For example, as shown in FIG. 5, an image CNN in FIG. 5 represents extracting a feature of an image block by using the image CNN. This application focuses on that for a given image, after the image is divided into image blocks, features of respective image blocks are extracted. It is assumed that there are five categories of image information in total, namely, people, grass, sky, dogs, and cats. Assuming that there are four image blocks for extraction, for each image block, a probability that the image block includes the foregoing five categories of image information. Then, for each category, a maximum probability is selected from the probabilities of the category of four image blocks. Maximum probabilities of the respective categories constitute the local feature of the image sample.

[0050] In an embodiment, alternatively, for each category, a maximum probability of the category may be compared with a corresponding probability threshold, to determine whether at least one image block in the sample image includes image information of the category. For example, for the category of people, a maximum probability is 0.99 in four image blocks, and a corresponding probability threshold is 0.8, indicating that at least one of the four image blocks includes a person.

[0051] 3) With regard to extraction of a global representation (global feature) and a local representation (local feature) of a text sample.

[0052] In an embodiment of this application, in order to better extract a global representation in a text sample, the following operations are included:

performing word segmentation on the text sample;

for each word segment, determining a vector of the word segment, different word segments having the same vector length; and

inputting vectors of word segments of the same text sample into a convolutional neural network to extract the global feature of the text sample, the neural network used to extract the global feature of the text sample including a plurality of convolutional layers and a pooling layer connected to the plurality of convolutional layers, and a field of view of a specified size of a previous convolutional layer being used as input of a current convolutional layer, the field of view of the specified size including features of vectors at least two word segments extracted by the previous convolutional layer.



[0053] For example, during a specific implementation, a convolutional operation of a convolutional layer may be represented by using the following formula (5):

where



[0054] In the formula (5), r represents a field of view of a specified size; l-1 represents the (l-1)th convolutional layer (that is, a previous convolutional layer); / represents the lth convolutional layer (that is, a current convolutional layer); Wl,f Wl,f represents a product matrix; frepresents the fth feature; i represents the ith location information; and

represents a feature extracted from the lth convolutional layer that corresponds to the fth feature of the ith location.

[0055] An operation of the pooling layer may be represented by the following formula (6):

where

represents output of the pooling layer; l+1 represents the pooling layer;

represents the (M×i)th feature extracted from the lth layer; and M represents a constant used to control a size of a pooling operation.

[0056] For example, as shown in FIG. 6, an inputted text is "a young child run through a large field of grass", a global representation of the text is extracted through a text convolutional neural network (that is, the sentence CNN in FIG. 6). Specifically, the sentence CNN includes a plurality of convolutional layers to extract features of the text. Then, a max-pooling layer is used to perform integration and dimension reduction on the features extracted by the convolutional layer to obtain a global representation of the text.

[0057] Accordingly, in this embodiment of this application, features of a text sample may be extracted through a plurality of convolutional layers, and then, useless feature information is filtered out through a pooling layer. Moreover, when a convolutional operation is performed by using the plurality of convolutional layers, a field of view of a specified size is used. Since the field of view includes features of a plurality of word segments, it is equivalent to learning organization and interaction information between adjacent word segments in a global representation of a text sample, so that a finally extracted global representation is more appropriate and accurate.

[0058] In an embodiment, a word2vec tool may be used to represent each word segment with a vector of a fixed length.

[0059] In addition, a neural network used to extract a global representation of a text may be a convolutional neural network or a recurrent neural network (RNN).

[0060] A local representation of the text sample may be extracted by using various methods, and can be more focused on each word segment. For example, a conventional term frequency-inverse document frequency (TF-IDF) may be used, or the Fisher Vector acts on a vector of each word segment or an average value of vectors of word segments. Such a representation is more focused on a word segment itself rather than considering relevance between word segments as a whole, and therefore, may be considered as a local representation of the text.

[0061] A image-text bidirectional search method based on the foregoing matching model in the embodiments of this application is described below.

[0062] As shown in FIG. 7, FIG. 7 is a schematic flowchart of a training method for an image-text matching model according to an embodiment of this application, the method includes the following steps.

[0063] Step 701: Extract a global representation of an image sample based on a global image representation CNN.

[0064] Step 702: Divide the image sample into a specified quantity of image blocks, and for each image block, calculate, based on a local image CNN, a probability that the image block includes a specified category of image information, select maximum probabilities of respective specified categories of image information in the specified quantity of image blocks, the maximum probabilities of the respective specified categories of image information constituting the local representation of the image sample.

[0065] Step 703: Perform word segmentation on a text sample, for each word segment, determine a vector of the word segment, different word segments having the same vector length, and input vectors of word segments of the same text sample into a pre-constructed convolutional neural network to extract a global representation of the text sample, the neural network used to extract the global representation of the text sample including a plurality of convolutional layers and a pooling layer connected to the plurality of convolutional layers, and a field of view of a specified size of a previous convolutional layer being used as input of a current convolutional layer, the field of view of the specified size including features of vectors at least two word segments extracted by the previous convolutional layer.

[0066] Step 704: Extract a local representation of the text sample.

[0067] The execution sequence of steps 701 to 704 is not limited to what is described in the above embodiments.

[0068] Step 705: Respectively map the respective global representations of the image sample and the text sample to a specified semantic space through at least two fully connected layers, and calculate a similarity between global representations of a heterogeneous sample pair (image-text pair) constituted by the image sample and the text sample, parameters of the fully connected layers being determined according to a preset target function.

[0069] Step 706: Respectively map the respective local representations of the image sample and the text sample to the specified semantic space through the at least two fully connected layers, and calculate a similarity between local representations of the heterogeneous sample pair (image-text pair), the parameters of the fully connected layers being determined according to the preset target function.

[0070] The execution sequence of steps 705 and 706 is not limited to what is described in the above embodiments.

[0071] In addition, the preset target function has been described in the foregoing embodiments, and is not described in detail herein again.

[0072] Step 707: Determine a matching degree of the heterogeneous sample pair using a weighted summation manner according to a preset weight of the similarity between the global representations and a preset weight of the similarity between the local representations.

[0073] In this embodiment of the present application, the local representation of the image is extracted through image blocks, and the global representation of the text is extracted by using a plurality of convolutional layers and a pooling layer, so that a relationship between word segments of the text can be obtained, and the extracted features are more accurate and comprehensive. Then, the preset target function is used to make a similarity of a semantically associated heterogeneous sample pair higher than a similarity of a non-semantically associated heterogeneous sample pair and a similarity between text sample associated with the same image sample higher than a similarity between text sample associated with different image samples, so that the calculation of a final similarity satisfies actual requirements better, so as to improve the accuracy of matching between the image and the text.

[0074] As shown in FIG. 8, FIG. 8 is a schematic flowchart of an image-text bidirectional search method based on the matching model described in the foregoing embodiments of this application. The method is carried out by a computer, the computer may be a server 12, and the method includes the following steps.

[0075] Step 801: Receive a reference sample, the reference sample being a text or an image.

[0076] Step 802: Extract a global feature and a local feature of the reference sample.

[0077] Step 803: Input the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample and a global feature and a local feature of the corresponding material.

[0078] A material library may be established to calculate a matching degree between the reference sample and each corresponding material in the material library. A method for calculating the matching degree has been described in the foregoing embodiments, and is not described in detail herein again.

[0079] Step 804: Select a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.

[0080] In this embodiment of this application, a corresponding material of a reference sample may be searched based on a matching model of a global representation and a local representation, so that a finally obtained matching result is more accurate. For example, an image matching a text is searched for or a question matching an image is searched for.

[0081] Based on the same inventive concept as that in the foregoing embodiments, an embodiment of this application further provides a training apparatus for an image-text matching model. The principle and the beneficial effects of the apparatus are similar to those of the foregoing method embodiments, and are not described in detail herein again.

[0082] FIG. 9 is a schematic diagram of a structure of the apparatus. The apparatus includes:

an image feature extraction module 901, configured to extract a global feature and a local feature of an image sample;

a text feature extraction module 902, configured to extract a global feature and a local feature of a text sample; and

a training module 903, configured to train a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an inputted image and a global feature and a local feature of an inputted text, a matching degree between the image and the text.



[0083] In some embodiments, the respective global features of the image and the text are mapped to a specified semantic space through the matching model, and a similarity between the global features of the image and the text is calculated; the respective local features of the image and the text are mapped to the specified semantic space, and a similarity between the local features of the image and the text is calculated; and the matching degree between the image and the text is determined using a weighted summation manner according to a preset weight of the similarity between the global features and a preset weight of the similarity between the local features.

[0084] In some embodiments, the matching model includes at least two fully connected layers, and the model parameters of the matching model include parameters of the at least two fully connected layers.

[0085] The training module 903 includes:

a global training unit 9031, configured to map the respective global features of the image sample and the text sample to the specified semantic space through the at least two fully connected layers; and

a local training unit 9032, configured to map the respective local features of the image sample and the text sample to the specified semantic space through the at least two fully connected layers.



[0086] The parameters of the at least two fully connected layers are determined according to mapping results and a preset target function, the preset target function being used to determine that a similarity of a semantically associated image-text pair is higher than a similarity of a non-semantically associated image-text pair; or
the preset target function being used to determine that a similarity of a semantically associated image-text pair is higher than a similarity of a non-semantically associated image-text pair, and that a similarity between text samples associated with the same image sample is higher than a similarity between text samples associated with different image samples.

[0087] In some embodiments, the preset target function is used to determine that a similarity of a semantically associated image-text pair is higher than a similarity of a non-semantically associated image-text pair.

[0088] The mapping results include: a global feature and a local feature of the image sample represented by the parameters of the at least two fully connected layers, and a global feature and a local feature of the text sample represented by the parameters of the at least two fully connected layers.

[0089] The training module 903 is configured to:

determine a similarity of a semantically associated image-text pair represented by the parameters of the at least two fully connected layers and a similarity of a non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers according to the global feature and the local feature of the image sample represented by the parameters of the at least two fully connected layers, and the global feature and the local feature of the text sample represented by the parameters of the at least two fully connected layers; and

input the similarity of the semantically associated image-text pair represented by the parameters of the at least two fully connected layers and the similarity of the non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers into the preset target function, to determine the parameters of the at least two fully connected layers.



[0090] If the preset target function is used to realize that a similarity of a semantically associated heterogeneous sample pair is higher than a similarity of a non-semantically associated heterogeneous sample pair, the preset target function may be:

where

represents the preset target function;

and

represent the text sample;

and

represent the image sample; d() represents a similarity between global representations of the text sample and the image sample when a similarity between global representations is determined; d()d() represents a similarity between local representations of the text sample and the image sample when a similarity between local representations is determined; λ1λ1 and λ2λ2 both represent preset coefficients;

represents the similarity of the semantically associated heterogeneous sample pair;

and

both represent the similarity of the non-semantically associated heterogeneous sample pair; and

µ1 and µ2 both represent preset thresholds.

[0091] In some embodiments, the preset target function is used to determine that a similarity of a semantically associated image-text pair is higher than a similarity of a non-semantically associated image-text pair, and that a similarity between text samples associated with the same image sample is higher than a similarity between text samples associated with different image samples.

[0092] The mapping results include: a global feature and a local feature of the image sample represented by the parameters of the at least two fully connected layers, and a global feature and a local feature of the text sample represented by the parameters of the at least two fully connected layers.

[0093] The training module 903 is configured to:

[0094] determine a similarity of a semantically associated image-text pair represented by the parameters of the at least two fully connected layers, a similarity of a non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers, a similarity between text samples associated with the same image sample represented by the parameters of the at least two fully connected layers, and a similarity between text samples associated with different image samples represented by the parameters of the at least two fully connected layers according to the global feature and the local feature of the image sample represented by the parameters of the at least two fully connected layers, and the global feature and the local feature of the text sample represented by the parameters of the at least two fully connected layers; and

[0095] input the similarity of the semantically associated image-text pair represented by the parameters of the at least two fully connected layers, the similarity of the non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers, the similarity between text samples associated with the same image sample represented by the parameters of the at least two fully connected layers, and the similarity between text samples associated with different image samples represented by the parameters of the at least two fully connected layers into the preset target function, to determine the parameters of the at least two fully connected layers.

[0096] In some embodiments, the preset target function is determined through one of the following:

and

where



[0097] LY is the preset target function,

represents a relationship between the similarity of the semantically associated image-text pair and the similarity of the non-semantically associated image-text pair, and L(Si, Sl, Sj) represents a relationship between the similarity between text samples associated with the same image sample and the similarity between text samples associated with different image samples;

and

represent the text sample;

and

represent the image sample; d()d() represents a similarity between global features of the text sample and the image sample when a similarity between global features is determined; d()d() represents a similarity between local features of the text sample and the image sample when a similarity between local features is determined; λ1λ1 and λ2λ2 both represent preset coefficients;

represents a similarity of a semantically associated heterogeneous sample pair (image-text pair);

and

both represent a similarity of a non-semantically associated heterogeneous sample pair (image-text pair); and

and u2 both represent preset thresholds, where

where Si, Sl Si, Sl represent the text samples semantically associated with the same image sample; Si, Sj Si, Sj represent the text samples associated with different image samples; d()d() represents the similarity between global features of the text samples when the similarity between global features is determined; d()d() represents the similarity between local features of the text samples when the similarity between local features is determined; and u3u3 represents a preset threshold.

[0098] In some embodiments, the image feature extraction module is configured to divide the image sample into a specified quantity of image blocks, and for each image block, calculate a probability that the image block includes a specified category of image information; and
select maximum probabilities of respective specified categories of image information in the specified quantity of image blocks, the maximum probabilities of the respective specified categories of image information constituting the local feature of the image sample.

[0099] In some embodiments, the image feature extraction module is configured to perform word segmentation on a text sample;
for each word segment, determine a vector of the word segment, different word segments having the same vector length, and
input vectors of word segments of the same text sample into a convolutional neural network to extract the global feature of the text sample, the neural network used to extract the global feature of the text sample including a plurality of convolutional layers and a pooling layer connected to the plurality of convolutional layers, and a field of view of a specified size of a previous convolutional layer being used as input of a current convolutional layer, the field of view of the specified size including features of vectors of at least two word segments extracted by the previous convolutional layer.

[0100] Based on the same inventive concept as that in the foregoing embodiments, an embodiment of this application further provides an image-text bidirectional search apparatus. The principle and the beneficial effects of the apparatus are similar to those of the foregoing method embodiments, and are not described in detail herein again.

[0101] FIG. 10 is a schematic diagram of a structure of the apparatus. The apparatus includes:

a reference sample receiving module 1001, configured to receive a reference sample, the reference sample being a text or an image;

a reference sample feature extraction module 1002, configured to extract a global feature and a local feature of the reference sample;

a search module 1003, configured to input the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample and a global feature and a local feature of the corresponding material; and

a selection module 1004, configured to select a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.



[0102] An embodiment of this application further provides a computing device. The computing device may be a desktop computer, a portable computer, a smartphone, a tablet computer, a personal digital assistant (PDA), or the like. As shown in FIG. 11, the computing device may include a central processing unit (CPU) 1101, a memory 1102, an input device 1103, an output device 1104, and the like. The input device may include a keyboard, a mouse, a touchscreen, and the like. The output device may include a display device such as a liquid crystal display (LCD) or a cathode ray tube (CRT).

[0103] The memory may include a read-only memory (ROM) and a random access memory (RAM), and provides program instructions and data stored in the memory to a processor. In this embodiment of this application, the memory may be configured to store program instructions of a training method for an image-text matching model and/or an image-text bidirectional search method. The processor invokes the program instructions stored in the memory, to perform the following operations according to the program instructions: extracting a global feature and a local feature of an image sample; extracting a global feature and a local feature of a text sample; and training a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an inputted image, and a global feature and a local feature of an inputted text, a matching degree between the image and the text.

[0104] Alternatively, the processor invokes the program instructions stored in the memory, to perform the following operations according to the program instructions: receiving a reference sample, the reference sample being a text or an image; extracting a global feature and a local feature of the reference sample; inputting the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample, and a global feature and a local feature of the corresponding material; and selecting a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.

[0105] This application provides a computer storage medium for storing computer program instructions used by the foregoing computing device, and including a program for performing the foregoing training method for an image-text matching model and/or the foregoing image-text bidirectional search method.

[0106] The computer storage medium may be any available medium or data storage device that is accessible by a computer, which includes, but is not limited to, a magnetic memory (for example, a floppy disk, a hard disk, a magnetic tape, and a magneto-optical disk (MO)), an optical memory (for example, a CD, a DVD, a BD, and an HVD), a semiconductor memory (for example, a ROM, an EPROM, an EEPROM, a non-volatile memory (NAND FLASH), and a solid state drive (SSD)), and the like.

[0107] Finally, the above embodiments are merely provided for describing the technical solutions of this application, but not intended to limit this application. It is to be understood by a person of ordinary skill in the art that although this application has been described in detail with reference to the foregoing embodiments, modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent replacements can be made to some technical features in the technical solutions, as long as such modifications or replacements do not cause the essence of corresponding technical solutions to depart from the spirit and scope of this application.


Claims

1. A training method for an image-text matching model, carried out by a computer and comprising:

extracting a global feature and a local feature of an image sample;

extracting a global feature and a local feature of a text sample; and

training a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an inputted image and a global feature and a local feature of an inputted text, a matching degree between the image and the text.


 
2. The method according to claim 1, further comprising:
mapping the respective global features of the image and the text to a specified semantic space through the matching model, and calculating a similarity between the global features of the image and the text; mapping the respective local features of the image and the text to the specified semantic space, and calculating a similarity between the local features of the image and the text; and determining the matching degree between the image and the text using a weighted summation manner according to a preset weight of the similarity between the global features and a preset weight of the similarity between the local features.
 
3. The method according to claim 1, wherein the matching model comprises at least two fully connected layers, and the model parameters of the matching model comprise parameters of the at least two fully connected layers; and
the training a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model comprises: mapping the respective global features of the image sample and the text sample to the specified semantic space through the at least two fully connected layers; mapping the respective local features of the image sample and the text sample to the specified semantic space through the at least two fully connected layers; and
determining the parameters of the at least two fully connected layers according to mapping results and a preset target function, the preset target function being used to determine that a similarity of a semantically associated image-text pair is higher than a similarity of a non-semantically associated image-text pair, or
the preset target function being used to determine that a similarity of a semantically associated image-text pair is higher than a similarity of a non-semantically associated image-text pair, and that a similarity between text samples associated with a same image sample is higher than a similarity between text samples associated with different image samples.
 
4. The method according to claim 3, wherein when the preset target function is used to determine that the similarity of the semantically associated image-text pair is higher than the similarity of the non-semantically associated image-text pair,
the mapping results comprise: a global feature and a local feature of the image sample represented by the parameters of the at least two fully connected layers, and a global feature and a local feature of the text sample represented by the parameters of the at least two fully connected layers; and
the determining the parameters of the at least two fully connected layers according to mapping results and a preset target function comprises:

determining a similarity of a semantically associated image-text pair represented by the parameters of the at least two fully connected layers and a similarity of a non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers according to the global feature and the local feature of the image sample represented by the parameters of the at least two fully connected layers, and the global feature and the local feature of the text sample represented by the parameters of the at least two fully connected layers; and

inputting the similarity of the semantically associated image-text pair represented by the parameters of the at least two fully connected layers and the similarity of the non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers into the preset target function, to determine the parameters of the at least two fully connected layers.


 
5. The method according to claim 4, wherein
the preset target function is:

wherein

represents the preset target function;

and

represent the text sample;

and

represent the image sample; d()d() represents a similarity between global features of the text sample and the image sample when a similarity between global features is determined; d()d() represents a similarity between local features of the text sample and the image sample when a similarity between local features is determined; λ1λ1 and λ2λ2 both represent preset coefficients;

represents the similarity of the semantically associated image-text pair;

and

both represent the similarity of the non-semantically associated image-text pair; and

and µ2 both represent preset thresholds.
 
6. The method according to claim 3, wherein when the preset target function is used to determine that the similarity of the semantically associated image-text pair is higher than the similarity of the non-semantically associated image-text pair, and that the similarity between text samples associated with the same image sample is higher than the similarity between text samples associated with different image samples,
the mapping results comprise: a global feature and a local feature of the image sample represented by the parameters of the at least two fully connected layers, and a global feature and a local feature of the text sample represented by the parameters of the at least two fully connected layers; and
the determining the parameters of the at least two fully connected layers according to mapping results and a preset target function comprises:

determining a similarity of a semantically associated image-text pair represented by the parameters of the at least two fully connected layers, a similarity of a non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers, a similarity between text samples associated with a same image sample represented by the parameters of the at least two fully connected layers, and a similarity between text samples associated with different image samples represented by the parameters of the at least two fully connected layers according to the global feature and the local feature of the image sample represented by the parameters of the at least two fully connected layers, and the global feature and the local feature of the text sample represented by the parameters of the at least two fully connected layers; and

inputting the similarity of the semantically associated image-text pair represented by the parameters of the at least two fully connected layers, the similarity of the non-semantically associated image-text pair represented by the parameters of the at least two fully connected layers, the similarity between text samples associated with the same image sample represented by the parameters of the at least two fully connected layers, and the similarity between text samples associated with different image samples represented by the parameters of the at least two fully connected layers into the preset target function, to determine the parameters of the at least two fully connected layers.


 
7. The method according to claim 6, wherein the preset target function is determined through one of following:

and

wherein

wherein LY is the preset target function,

represents a relationship between the similarity of the semantically associated image-text pair and the similarity of the non-semantically associated image-text pair, and L(Si, Sl, Sj) represents a relationship between the similarity between text samples associated with the same image sample and the similarity between text samples associated with different image samples;

and

represent the text sample;

and

represent the image sample; d()d() represents a similarity between global features of the text sample and the image sample when a similarity between global features is determined; d()d() represents a similarity between local features of the text sample and the image sample when a similarity between local features is determined; λ1λ1 λ2 and λ2 both represent preset coefficients;

represents the similarity of the semantically associated image-text pair;

and

both represent the similarity of the non-semantically associated image-text pair;

and u2 both represent preset thresholds; and

wherein Si, Sl Si, Sl represent the text samples semantically associated with the same image sample; Si, Sj Si, Sj represent the text samples associated with different image samples; d()d() represents the similarity between global features of the text samples when the similarity between global features is determined; d()d() represents the similarity between local features of the text samples when the similarity between local features is determined; and u3u3 represents a preset threshold.
 
8. The method according to claim 1, wherein the extracting a local feature of an image sample comprises:

dividing the image sample into a specified quantity of image blocks, and for each image block, calculating a probability that the image block comprises a specified category of image information; and

selecting maximum probabilities of respective specified categories of image information in the specified quantity of image blocks, the maximum probabilities of the respective specified categories of image information constituting the local feature of the image sample.


 
9. The method according to claim 1, wherein the extracting a global feature of a text sample comprises:

performing word segmentation on the text sample;

for each word segment, determining a vector of the word segment, different word segments having the same vector length; and

inputting vectors of word segments of a same text sample into a convolutional neural network to extract the global feature of the text sample, the neural network used to extract the global feature of the text sample comprising a plurality of convolutional layers and a pooling layer connected to the plurality of convolutional layers, and a field of view of a specified size of a previous convolutional layer being used as input of a current convolutional layer, the field of view of the specified size comprising features of vectors of at least two word segments extracted by the previous convolutional layer.


 
10. An image-text bidirectional search method, carried out by a computer and comprising:

receiving a reference sample, the reference sample being a text or an image;

extracting a global feature and a local feature of the reference sample;

inputting the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample and a global feature and a local feature of the corresponding material; and

selecting a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.


 
11. A training apparatus for an image-text matching model, comprising:

an image feature extraction module, configured to extract a global feature and a local feature of an image sample;

a text feature extraction module, configured to extract a global feature and a local feature of a text sample; and

a training module, configured to train a matching model according to the extracted global feature and local feature of the image sample and the extracted global feature and local feature of the text sample, to determine model parameters of the matching model, the matching model being used to determine, according to a global feature and a local feature of an image and a global feature and a local feature of a text, a matching degree between the image and the text.


 
12. An image-text bidirectional search apparatus, comprising:

a reference sample receiving module, configured to receive a reference sample, the reference sample being a text or an image;

a reference sample feature extraction module, configured to extract a global feature and a local feature of the reference sample;

a search module, configured to input the global feature and the local feature of the reference sample into a matching model, to enable the matching model to calculate a matching degree between the reference sample and a corresponding material; the corresponding material being an image when the reference sample is a text, and the corresponding material being a text when the reference sample is an image; and the matching model being capable of determining the matching degree between the reference sample and the corresponding material based on the global feature and the local feature of the reference sample and a global feature and a local feature of the corresponding material; and

a selection module, configured to select a corresponding material whose matching degree is greater than a specified matching degree as a material matching the reference sample.


 
13. A computing device, comprising a memory and a processor, the memory being configured to store program instructions, and the processor being configured to invoke the program instructions stored in the memory, to perform the training method for an image-text matching model according to any one of claims 1 to 9 according to the program instructions.
 
14. A computer storage medium, storing computer-executable instructions, the computer-executable instructions being used to enable the computer to perform the training method for an image-text matching model according to any one of claims 1 to 9 according to the computer-executable instructions.
 
15. A computing device, comprising a memory and a processor, the memory being configured to store program instructions, and the processor being configured to invoke the program instructions stored in the memory to perform the image-text bidirectional search method according to claim 10 according to the program instructions.
 
16. A computer storage medium, storing computer-executable instructions, the computer-executable instructions being used to enable the computer to perform the image-text bidirectional search method according to claim 10 according to the computer-executable instructions.
 




Drawing






















Search report










Cited references

REFERENCES CITED IN THE DESCRIPTION



This list of references cited by the applicant is for the reader's convenience only. It does not form part of the European patent document. Even though great care has been taken in compiling the references, errors or omissions cannot be excluded and the EPO disclaims all liability in this regard.

Patent documents cited in the description