WO2021184776A1 - 图像识别方法、装置、计算机设备和存储介质 - Google Patents

图像识别方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2021184776A1
WO2021184776A1 PCT/CN2020/127455 CN2020127455W WO2021184776A1 WO 2021184776 A1 WO2021184776 A1 WO 2021184776A1 CN 2020127455 W CN2020127455 W CN 2020127455W WO 2021184776 A1 WO2021184776 A1 WO 2021184776A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
image recognition
label
tag
concept
Prior art date
Application number
PCT/CN2020/127455
Other languages
English (en)
French (fr)
Inventor
王汉杰
李岩
毛懿荣
成文龙
谌丹璐
陈波
Original Assignee
腾讯科技（深圳）有限公司
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 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2021184776A1 publication Critical patent/WO2021184776A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches

Definitions

  • This application relates to the field of image recognition technology, in particular to an image recognition method, device, computer equipment and storage medium.
  • Image recognition technology refers to a technology that recognizes objects included in an image, such as identifying whether an image includes a cow.
  • image recognition can be performed through an image recognition model based on artificial intelligence.
  • the image recognition model needs to be trained in advance based on the training data.
  • the image tags that need to be output in different business scenarios are often different. Therefore, when image recognition is needed, model training needs to be performed according to the image tags that need to be recognized, and then image recognition is performed based on the image recognition model obtained by training. In this case, computer resources are wasted and the efficiency of image recognition is low.
  • an image recognition method is provided.
  • An image recognition method executed by a computer device, the method comprising: receiving an image recognition request for a target concept tag; obtaining a target image to be recognized corresponding to the image recognition request; and inputting the target image into an entity tag corresponding Image recognition is performed in the target image recognition model of, to obtain a target image recognition result set, the image recognition result in the target image recognition result set is the image recognition result corresponding to the target entity tag, and the target entity tag is the target concept tag Corresponding entity tag; the image recognition result corresponding to the target concept tag is obtained according to the target image recognition result set.
  • An image recognition device comprising: an image recognition request receiving module for receiving an image recognition request for a target concept label; a target image obtaining module for obtaining a target image to be recognized corresponding to the image recognition request;
  • the target image recognition result set acquisition module is used to input the target image into the target image recognition model corresponding to the entity tag for image recognition to obtain a target image recognition result set, and the image recognition result in the target image recognition result set is The image recognition result corresponding to the target entity tag, where the target entity tag is the entity tag corresponding to the target concept tag; the result obtaining module is configured to obtain the image recognition result corresponding to the target concept tag according to the target image recognition result set .
  • a computer device includes a memory and a processor, wherein computer readable instructions are stored in the memory, and when the computer readable instructions are executed by the processor, the following steps are implemented: receiving an image recognition request for a target concept label; acquiring The target image to be recognized corresponding to the image recognition request; the target image is input into the target image recognition model corresponding to the entity tag for image recognition, and the target image recognition result set is obtained, and the image in the target image recognition result set
  • the recognition result is an image recognition result corresponding to a target entity tag, and the target entity tag is an entity tag corresponding to the target concept tag; the image recognition result corresponding to the target concept tag is obtained according to the target image recognition result set.
  • One or more non-volatile storage media storing computer-readable instructions, which, when executed by one or more processors, implement the following steps: receive an image recognition request for the target concept label; obtain the The target image to be recognized corresponding to the image recognition request; the target image is input into the target image recognition model corresponding to the entity tag for image recognition, and the target image recognition result set is obtained, and the image recognition result in the target image recognition result set Is the image recognition result corresponding to the target entity tag, the target entity tag is the entity tag corresponding to the target concept tag; the image recognition result corresponding to the target concept tag is obtained according to the target image recognition result set.
  • FIG. 1 is an application environment diagram of image recognition methods in some embodiments
  • Figure 2 is a schematic flowchart of an image recognition method in some embodiments
  • Figure 3 is a schematic diagram of an image tag tree in some embodiments.
  • FIG. 4 is a schematic diagram of an interface for triggering an image recognition request and displaying an image recognition result corresponding to a target concept label in other embodiments;
  • FIG. 5 is a schematic diagram of a flow of inputting a target image into a target image recognition model corresponding to an entity tag for image recognition in some embodiments to obtain a target image recognition result set;
  • FIG. 6 is a flowchart of performing model training on an image recognition model to obtain a target image recognition model in some embodiments
  • Figure 7 is a schematic diagram of a label modification interface in some embodiments.
  • Figure 8 is a schematic diagram of an image recognition service customization interface in some embodiments.
  • Figure 9 is a schematic diagram of an interface for requesting users to collaborate in some embodiments.
  • FIG. 10 is a schematic diagram of an interface for a collaborative terminal to determine a collaborative task in some embodiments.
  • Figure 11 is a schematic diagram of a collaboration interface in some embodiments.
  • Figure 12 is a schematic diagram of the operation of the image recognition system in some embodiments.
  • Figure 13 is a structural block diagram of an image recognition device in some embodiments.
  • Figure 14 is a diagram of the internal structure of a computer device in some embodiments.
  • AI Artificial Intelligence
  • digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.
  • artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a similar way to human intelligence.
  • Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.
  • Artificial intelligence technology is a comprehensive discipline, covering a wide range of fields, including both hardware-level technology and software-level technology.
  • Basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, and mechatronics.
  • Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning/deep learning.
  • Computer Vision is a science that studies how to make machines "see”. Furthermore, it refers to the use of cameras and computers instead of human eyes to identify, track, and measure targets. And further graphics processing, so that computer processing becomes more suitable for human eyes to observe or send to the instrument to detect the image.
  • Computer vision studies related theories and technologies trying to establish an artificial intelligence system that can obtain information from images or multi-dimensional data.
  • Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content/behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, synchronous positioning and mapping Construction and other technologies also include common face recognition, fingerprint recognition and other biometric recognition technologies.
  • Machine Learning is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. Specializing in the study of how computers simulate or realize human learning behaviors in order to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve its own performance.
  • Machine learning is the core of artificial intelligence, the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence.
  • Machine learning and deep learning usually include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and style teaching learning.
  • artificial intelligence technology has been researched and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, autonomous driving, drones , Robotics, intelligent medical care, intelligent customer service, etc., I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play more and more important values.
  • the image recognition method provided in this application can be applied to the application environment as shown in FIG. 1.
  • the terminal 102 communicates with the server 104 through the network.
  • the server 104 is pre-deployed with a target image recognition model obtained by performing model training in advance.
  • the user can trigger an image recognition request for the target concept label by operating the terminal 102, the terminal 102 sends the image recognition request for the target concept label to the server 104, and the server 104 executes the method provided in the embodiment of the application to obtain The image recognition result corresponding to the target concept label.
  • the server 104 returns the image recognition result corresponding to the target concept tag to the terminal 102, and the terminal 102 displays the image recognition result corresponding to the target concept tag.
  • the image recognition method provided in the embodiment of the present application may also be applied to the terminal 102.
  • the target image recognition model may be deployed in the terminal 102.
  • the terminal 102 may be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server 104 may be implemented by an independent server or a server cluster composed of multiple servers.
  • an image recognition method is provided. Taking the method applied to the server in FIG. 1 as an example for description, the method includes the following steps:
  • Step S202 Receive an image recognition request for the target concept label.
  • the concept label is used to represent the concept
  • the entity label is used to represent the entity.
  • Concepts are the expressions of entities with common characteristics. Entity refers to the concrete thing corresponding to the concept.
  • a concept includes each entity it defines.
  • a concept tag corresponds to multiple entity tags. "Multiple" means "at least two".
  • the concept label can be a car, and the physical label can include a car of brand A and a car of brand B.
  • the concept tag may be a dog, and the entity tag may include a shepherd dog, a hunting dog, and so on.
  • the concept tag and the entity tag are relative, and whether an image tag is a concept tag or an entity tag can be changed according to actual conditions, such as application scenarios.
  • the former concept includes the meaning expressed by the latter concept
  • the former concept is a concept label relative to the latter concept.
  • the latter concept is an entity tag relative to the previous concept.
  • the "car” tag is an entity tag.
  • the "Transportation” tag is a concept tag.
  • the "car” tag is a concept tag, and the brands of cars such as "brand A" and "brand B" are physical tags.
  • the image recognition request is used to request the recognition of the content of the image.
  • the image recognition request for the target concept tag refers to: the image recognition request is used to request the content corresponding to the target concept tag in the recognition image.
  • an image recognition request for the "car" tag is used to request to recognize whether the image includes a car.
  • the target concept label may be one or more
  • the image recognition request may carry label related information corresponding to the target concept label
  • the label related information may include the concept label identification or the target service identification.
  • the server receives the image recognition request carrying the concept tag identifier, it obtains the tag corresponding to the concept tag identifier as the target concept tag, and determines that the image recognition request is an image recognition request for the target concept tag. It is also possible to pre-establish the corresponding relationship between the service ID and the concept label.
  • the server can determine the target concept label according to the corresponding relationship, and determine that the image recognition request is an image recognition request for the target concept label .
  • the server may pre-store the hierarchical relationship between the image tags, and the type of the image tag may be determined according to the hierarchy of the image tag, and the type of an image tag may be a concept tag or an entity tag.
  • the image tag with the lowest level can be used as the entity tag
  • the image tag of the upper level corresponding to the entity tag can be used as the concept tag.
  • the image label corresponding to the training image used for model training may be used as the entity label, and the label representing the upper concept corresponding to the entity may be the concept label.
  • the image label corresponding to the training image refers to the entity that directly corresponds to the image label included in the training image.
  • FIG. 3 it is a schematic diagram of an image tag tree in some embodiments.
  • the image tags can be displayed in a tree structure according to the hierarchical relationship, including non-leaf nodes and leaf nodes.
  • a leaf node refers to a node without child nodes, that is, the lowest-level node, such as the C1, C2, C3, and B2 nodes in FIG. 3.
  • the label of a non-leaf node represents a coarse or fine concept, so it is called a "concept label”.
  • the label of a leaf node represents a specific label corresponding to the training image, so it is called an "entity label".
  • the training image corresponding to the entity tag can be obtained to train the model.
  • Step S204 Acquire a target image to be recognized corresponding to the image recognition request.
  • the image recognition request may carry the image or image index information.
  • the image recognition request When the image recognition request carries an image, the image is used as the target image to be recognized.
  • the image index information may be, for example, the storage location of the image or the identification such as the name of the image.
  • the image recognition request When the image recognition request carries image index information, the corresponding image can be obtained according to the image index information as the target image to be recognized.
  • the "acc" folder stores 5 pictures.
  • the image recognition request can carry the name of the folder "acc”, and the server can add "acc"
  • the 5 pictures in the "acc” folder are used as the target images to be recognized.
  • Step S206 Input the target image into the target image recognition model corresponding to the entity tag for image recognition to obtain a target image recognition result set.
  • the image recognition result in the target image recognition result set is the image recognition result corresponding to the target entity tag, and the target entity
  • the label is the entity label corresponding to the target concept label.
  • the image recognition model is used to recognize the image, and the image recognition model may be, for example, a convolutional neural network model.
  • the target image recognition model corresponding to the entity tag refers to: the target image recognition model outputs the recognition result corresponding to the entity tag.
  • the target entity tag is the entity tag corresponding to the target concept tag, and the entity corresponding to the target entity tag is the entity belonging to the target concept. For example, if the target concept label is "computer equipment”, the target entity label can be "tablet” and "server”.
  • the image recognition result corresponding to the target entity tag indicates that the target image includes the entity corresponding to the target entity tag.
  • the image recognition result corresponding to the target entity tag can be the likelihood that the target image output by the model includes the entity corresponding to the entity tag, where the likelihood refers to the likelihood that the entity is included in the image, and the greater the likelihood that the entity tag corresponds, it means The more likely there is an entity corresponding to the entity tag in the image, the probability can be expressed by probability.
  • the target image recognition model may be obtained through machine learning training in advance, and a supervised training method may be used to obtain training images and corresponding entity tags for model training.
  • the goal of model training can be to make the difference between the probability distribution corresponding to the entity label output by the model and the probability distribution corresponding to the entity label corresponding to the training image as small as possible. That is, the parameters of the model can be adjusted in the direction of decreasing the loss value of the model. Until the model converges, the target image recognition model is obtained.
  • the difference between the probability distribution corresponding to the entity label output by the model and the probability distribution corresponding to the entity label corresponding to the training image can be positively correlated with the model loss value, that is, the greater the difference, the greater the model loss value.
  • the server can input the target image into the target image recognition model, and the target image recognition model can process the image, for example, through the processing of the convolutional layer, the fully connected layer, and the activation layer, and output each entity label Corresponding probability, the probability corresponding to the target entity tag can be obtained to form a target image recognition result set.
  • Step S208 Obtain an image recognition result corresponding to the target concept tag according to the target image recognition result set.
  • the image recognition result corresponding to the target concept label may be at least one of the likelihood corresponding to the target concept label and the judgment result of the entity corresponding to the target concept label. For example, when the target concept tag is "transportation”, the image recognition result corresponding to the target concept tag may be "the probability of including a transportation in the image is 0.9", or the judgment result "includes a transportation in the image”.
  • the server Since the image recognition result in the target image recognition result set is the recognition result corresponding to the target entity tag, and the concept represented by the target concept tag includes the entity represented by the target entity tag, the server The image recognition result corresponding to the target concept label can be obtained according to the target image recognition result set.
  • the target image recognition model is a recognition model that recognizes whether the image includes entities such as “car”, “aircraft”, or “ship”.
  • the target image is input into the target image recognition model, and the probabilities corresponding to the output "car”, “aircraft”, and “ship” are “0.9”, “0.05”, and “0.05” respectively.
  • the image recognition result corresponding to the target entity tag can be "0.9”, “0.05”, and “0.05”. Since the maximum probability of “0.9” corresponding to “car” is obtained according to “0.9”, “0.05”, and “0.05”, which is greater than the preset probability, such as 0.8, the image recognition result corresponding to the target concept label can be that the target image includes “ Transportation”.
  • the image to be recognized can be input into the target image recognition model corresponding to the entity tag for image recognition, and the target entity tag is obtained According to the corresponding image recognition result, the image recognition result corresponding to the target concept tag is obtained according to the image recognition result corresponding to the target entity tag. Therefore, when image recognition is required for the target concept label, no additional model training is required for the target concept label, and the image recognition result corresponding to the target concept label can also be obtained, which reduces the occupation of computer resources and improves the efficiency of image recognition.
  • the image recognition result corresponding to the target entity tag includes the possibility that the target image includes the entity corresponding to the target entity tag
  • obtaining the image recognition result corresponding to the concept tag according to the target image recognition result set includes: recognizing from the target image Obtain the possibility degree that satisfies the possibility degree condition in the result set as the target possibility degree; obtain the image recognition result corresponding to the target concept label according to the target possibility degree.
  • the possibility degree condition includes the possibility degree sorting before the first sorting, and the possibility degree sorting is the first At least one of the sorting or the possibility degree is greater than the first preset threshold.
  • the first ranking and the first preset threshold may be set according to needs, for example, the possibility condition may be the maximum possibility, or the possibility greater than 0.8.
  • the likelihood corresponding to the target concept label can be calculated according to the target likelihood, or it can be the judgment result of determining the entity corresponding to the target concept label according to the target possibility. For example, if the probability condition includes that the probability is greater than the first preset threshold, if the probability greater than the first preset threshold is greater than the preset number, for example, 3, it means that the possibility that the entity corresponding to the target concept tag exists in the image is If it is larger, it can be determined that the image recognition result corresponding to the target concept label is: the target image includes the entity corresponding to the target concept label.
  • the possibility condition may be the maximum possibility
  • obtaining the image recognition result corresponding to the target concept label according to the target possibility includes at least one of the following steps: when the target possibility is greater than a second preset threshold, the target concept is determined
  • the image recognition result corresponding to the label includes the entity corresponding to the target concept label included in the target image; or the target possibility is used as the image recognition result corresponding to the target concept label.
  • the second preset threshold may also be set as required. For example, it can be 0.85.
  • the maximum probability can be used as the image recognition result corresponding to the target concept label, or it can be determined that the target image includes the entity corresponding to the target concept label when it is determined that the target probability is greater than the second preset threshold.
  • the target concept label is "vehicle”
  • the target image recognition result set includes the output "car”, "aircraft”, and "ship” respectively.
  • the corresponding probabilities are "0.9", “0.05", and "0.05".
  • the maximum probability of "0.9” can be used as the probability of including "vehicles” in the target image. It can also be determined that when the maximum probability "0.9” is greater than the second preset probability "0.85", it is determined that a vehicle is included in the target image.
  • the image result may be returned to the target service terminal corresponding to the image recognition request sent.
  • FIG. 4 it is a schematic diagram of an interface for triggering an image recognition request and displaying an image recognition result corresponding to a target concept tag in some embodiments.
  • the interface includes a picture upload area 402, a judgment result display area 404, and a probability display area 406.
  • "XX business classifier" is the business name of the image recognition business. When the user needs to recognize the image, he can click the "Upload” button to enter the image upload interface to select the picture.
  • the target service terminal can trigger the server to send an image recognition request for the target concept label
  • the target concept tags can be, for example, "birds” and “fish”. Entity tags corresponding to output "birds” are deployed in the server, such as “sparrow”, “geese”, “partridge”, etc., and entity tags corresponding to "fish” such as “goldfish”, "carp", and "grass carp” are deployed.
  • Image recognition model based on the probability of image tags.
  • the server inputs the pictures uploaded by the terminal into the image recognition model.
  • the images uploaded by the image recognition model include “sparrow”, “geese”, “partridge”, “goldfish”, The probability of "carp” and "grass carp".
  • the server may obtain the maximum probability of the probabilities corresponding to the entity tags "Sparrow”, “Goose”, and “Partridge” as the probability of including the target concept tag "Bird” in the uploaded picture.
  • the server may obtain the maximum probability of the probabilities corresponding to the entity tags "goldfish”, “carp” and “grass carp” as the probability that the target concept tag "fish” is included in the uploaded picture. If the probability of "birds" is 0.9667, which is greater than the preset probability of 0.85, the server can also output the image recognition result of "birds in the picture".
  • step S206 is to input the target image into the target image recognition model corresponding to the entity tag for image recognition, and obtain the target image recognition result set includes the following steps:
  • Step S502 Input the target image into the target image recognition model corresponding to the entity tag for image recognition, and obtain a candidate image recognition result set.
  • the candidate image recognition result set includes the image recognition result corresponding to each first entity tag.
  • the first entity tag may be any entity tag.
  • the target image recognition model may be a general image recognition model, which may be used to output the probability of entity tags corresponding to multiple concept tags included in the output image.
  • different service parties have different image recognition requirements, and some service parties' services are related to games, so the image tags that need to be recognized are related to games.
  • Some business parties business is related to plants, and the image tags that need to be identified are related to plants.
  • a model that can recognize image tags related to plants and a general image recognition model that can recognize image tags related to games can be deployed.
  • the target image can be input into the target image recognition model.
  • the target image recognition model can process the image and output the probability corresponding to each first entity tag.
  • the server can obtain the probability corresponding to the first entity tag to form a candidate image recognition model. Result collection.
  • Step S504 Determine the first entity tag corresponding to the target concept tag as the target entity tag.
  • the corresponding relationship between the concept tag and the entity tag can be set in advance, and after the target concept tag is obtained, the entity tag corresponding to the target concept tag is obtained as the target entity tag.
  • the target concept tag is "computer equipment” and the first entity tag includes “sparrow”, “geese”, “partridge”, “goldfish”, “carp”, “grass carp”, “tablet” and “server”, then
  • the target entity tags corresponding to the target concept tag "computer equipment” can be “tablet” and "server”.
  • Step S506 Obtain the image recognition result corresponding to the target entity tag from the candidate image recognition result set to obtain the target image recognition result set.
  • the server after the server obtains the target entity tag, it can obtain the image recognition result corresponding to the target entity tag from the candidate image recognition result set to form the target image recognition result set.
  • the target image by inputting the target image into the target image recognition model corresponding to the entity tag for image recognition, a candidate image recognition result set composed of the image recognition results corresponding to each first entity tag is obtained, and the candidate image recognition result collection is
  • the image recognition result corresponding to the target entity tag is obtained in the, so for any target concept tag, it can be input into the same image recognition model, that is, the general recognition model for recognition, so there is no need to train a special entity tag for each concept tag.
  • the image recognition model saves model training resources.
  • FIG. 6 it is a flowchart of performing model training on an image recognition model in some embodiments to obtain a target image recognition model, including the following steps:
  • Step S602 Obtain a first entity tag set.
  • the first entity tag set includes entity tags corresponding to multiple concept tags.
  • the first entity tag set may be preset, may be manually set by a business party with image recognition requirements, or may be set by a service party that provides image recognition services.
  • the service party and the business party can log in and operate on the platform with their own identities.
  • the service party can organize the tags and publish the tags in the "tag store”.
  • the business party can also add tags as needed. It can be understood that the multiple concept tags include target concept tags.
  • the service party can collaborate with multiple people in the process of sorting and publishing labels, and experts in various fields are responsible for the labels in their respective fields, which not only reduces the workload, but also ensures professionalism.
  • the name of the server author who added or modified the tag can appear in the cursor floating window for other collaborators to view, ensuring that the responsibility is clear.
  • the number of tags can be set according to the actual situation, for example, it can reach tens of thousands or even hundreds of thousands.
  • the tags can display the hierarchical relationship in a tree structure. The hierarchical relationship helps to clarify the system context, quickly locate the tags, and facilitate the sorting and publishing of tags.
  • the tags sorted and published by the server may include concept tags and entity tags, and the server may receive a modification request to modify the concept tags and entity tags.
  • the type of modification may include at least one of adding tags, deleting tags, or changing tag information.
  • Adding tags refers to adding concept tags or entity tags to the specified parent node.
  • the server can delete the label according to the user's deletion operation, or restrict the deletion of the label, for example, it is set that the image label of the preset root node cannot be deleted.
  • the terminal can pop up a query box to ask whether to delete the label. After receiving the confirmation operation, delete it again to avoid misoperation.
  • Changing the label information can be to modify the description of the image label, or to modify the parent node of the label.
  • FIG. 7 it is a schematic diagram of a label modification interface in some embodiments.
  • the label modification interface can pop up, and the interface displays the functions of "add child node”, “update node”, “delete node” and "copy this label code” Control, by triggering the corresponding control, the corresponding function can be realized.
  • the types of label nodes can be represented by colors, and label nodes of different colors have different types.
  • the common node is gray, indicating that the label is a newly-added label, and the corresponding image has not been obtained.
  • the yellow node indicates that it supports viewing the image data of the label
  • the blue node supports viewing the image
  • the label is an entity tag.
  • the model has been trained based on the image corresponding to the entity tag, so the image recognition model can support the corresponding output entity tag The probability.
  • tags that support image viewing when a click operation on the tag is received, an image display interface can be displayed on the terminal, and the image and the annotation result of the image can be displayed on the interface.
  • step S604 an image corresponding to each first entity tag in the first entity tag set is obtained as a training image to obtain a training image set.
  • the image corresponding to the first entity tag refers to the image including the entity represented by the first entity tag.
  • the first entity tag includes "Sparrow” and "Mobile”. Then, images including sparrows and images including mobile phones can be obtained, and these images are used as training images to obtain a training image set.
  • Step S606 Perform model training according to the training image set to obtain a target image recognition model, and use the target image recognition model as a general image recognition model corresponding to multiple concept tags.
  • taking the target image recognition model as a general image recognition model corresponding to multiple concept tags means that the target image recognition model is universal, and for an image recognition request for any one of the multiple concept tags, the corresponding image recognition model can be obtained.
  • the target image is input into the general image recognition model for image recognition.
  • multiple concept tags include "birds" tags and "dogs" tags.
  • After receiving an image recognition request for the "birds” tag input the corresponding target image into the general image recognition model for image recognition.
  • Recognition After receiving the image recognition request for the "dog” tag, the corresponding target image can be input into the general image recognition model for image recognition.
  • the server after the server obtains the training image collection, it can input the training images in the training image collection into the image recognition model to be trained, and the model outputs the probability that the training image includes the entity corresponding to each entity label, which can be based on the probability of the model output Distribution, and the difference in the probability distribution of the real image label corresponding to the training image, the model loss value is obtained, and the stochastic gradient descent method is used to adjust the parameters of the model until the model converges, and the target image recognition model is obtained.
  • Each concept label corresponds to the entity label corresponding to the image training, so the target image recognition model can be used as a general image recognition model corresponding to multiple concept labels. When an image recognition request for these concept labels is received, you can enter Go to this universal image recognition model for image recognition.
  • an image recognition model for recognizing the entities represented by the entity tags of the multiple concept tags can be trained.
  • Use the image recognition model as a general image recognition model corresponding to multiple concept tags.
  • the same model can also be used for image recognition to obtain the recognition result of the entity tag, and then according to the entity tag
  • the recognition result obtains the image recognition result corresponding to the target concept label, so the image recognition service requirements of different service parties can be met without the need to train corresponding models for different service parties.
  • the image recognition service requirement of the business party A is bird image recognition service
  • the image recognition service requirement of service party B is the dog image recognition service.
  • it can be trained to obtain a general image recognition model that can recognize the entity tags corresponding to the bird image recognition service and the probabilities corresponding to the entity tags corresponding to the dog image recognition service.
  • the image recognition request sent can obtain the target image corresponding to the image recognition request, and input it into the general image recognition model for image recognition.
  • the images corresponding to the first entity tags may be pre-stored, or obtained through image mining, or offline image mining.
  • a set of candidate images and text corresponding to the images can be obtained, and the set of candidate images can be images obtained by crawling on the Internet.
  • the server can automatically obtain the training image set and train the model, so that the target image recognition model can potentially evolve with the accumulation of image data, and can use the evolved target image recognition model to update the old target image recognition in the image recognition engine Model to provide better services for the business side.
  • the image recognition engine refers to the core component service that takes the image as input, after model inference, and outputs the image label.
  • the server may send a tag color change instruction to change the color of the tags in the tag tree. For example, gray leaf nodes (with no corresponding image entity tag) can be changed to yellow (with corresponding image entity tag). , So that the business party can determine whether the label has corresponding image data according to the color of the label.
  • the server can also activate the image viewing function corresponding to the entity label according to the label color change instruction, for example, the "view image" function control corresponding to the entity label, after receiving the label color change instruction, can change from gray (indicating non-clickable) to black ( Means clickable).
  • the image recognition method may further include: receiving an image recognition service customization request sent by the target service terminal, the image recognition service customization request carrying the target service identifier and the target concept label; establishing the correspondence between the target service identifier and the target concept label . Therefore, when an image recognition request carrying a target service identifier is received, the image recognition request can be determined as an image recognition request for the target concept label according to the corresponding relationship.
  • the service ID can be used to identify the corresponding service, and the service ID can be set as required, can be automatically generated by the server, or can be obtained by inputting the service ID of the service party.
  • the target business party needs to customize the image recognition service of whether the recognition image includes "transportation”
  • you can enter the "transportation classifier” when the target business terminal logged in by the target business party receives the target business party to trigger the image customization service
  • the image recognition service customization request is triggered, and the target service identifier "Vehicle Classifier" is carried.
  • the image recognition service customization request is used to request the customization of the image recognition service corresponding to the "transportation tool classifier".
  • the server When receiving an image recognition service customization request, the server can store the target service ID and the target concept label in association. In this way, when an image recognition request carrying the target service ID is received, the server determines that the image recognition request is a target concept label according to the corresponding relationship. Image recognition request. Therefore, the user does not need to select the target concept tag every time image recognition is required, which improves the efficiency of image recognition.
  • the service customization portal "Customize my image recognition service” can be displayed on the target service terminal.
  • a trigger operation such as a click operation on the service customization portal
  • it can enter the image recognition service customization interface, as shown in Figure 8, Image Recognition
  • the service customization interface may include an image tag tree display area 802, a concept tag display area 804, and a service identification display area 806.
  • the image tag tree display area 802 is used to display candidate image tags.
  • the candidate image tags are displayed in the form of a tag tree.
  • the target service terminal can receive a user's tag selection operation, such as an operation of clicking a tag.
  • the label that can be selected by the target business party can be set as a label of a non-leaf node, that is, a conceptual label.
  • the terminal adds the label selected by the user to the concept label display area 804 for display.
  • the service identifier display area 806 displays the service identifier currently selected by the user. For example, “dog classifier” is the service identifier, indicating that the service is to classify dogs Business. If the user needs to create another classifier corresponding to another business, he can click the control "Create a new shopping cart” to create a new business.
  • the target business terminal receives the selection operation of the "Create a new shopping cart" control, the target business The terminal may display a service identification input box, and the target service terminal obtains the service identification entered by the user in the service identification input box as a new service identification.
  • “shopping cart” represents a collection of conceptual tags selected by the user.
  • the target service terminal corresponding to the target service party receives a trigger operation on the "training" control, such as a click operation or a voice operation, it can send an image recognition service customization request to the server, and the server generates the image recognition corresponding to the image recognition service customization request engine.
  • the target service terminal can also receive the selection operation of the "refresh” control.
  • the server refreshes the selected label corresponding to the service identifier according to the selection operation of the "refresh” control, and the target service terminal can receive the selection operation of the "preview” control.
  • the selection operation is an operation for selecting target information, and may be an operation to select information through one or more of a control, a voice, a gesture, or an expression.
  • the server may also receive a label query request sent by the target service terminal corresponding to the target business party.
  • the query request may carry query information.
  • the query information may be, for example, label name information or code.
  • the label of is returned to the target service terminal, so that the target service party can quickly query the label.
  • the image recognition platform can provide the functions of precise search and fuzzy search of image labels, so that business parties with a clear label list can quickly search for labels and speed up the addition of label shopping carts.
  • the target business party can query the label through the image label name and the image label code.
  • the target image recognition model may be obtained by performing model training after receiving an image recognition service customization request.
  • the target image recognition model can also be obtained through pre-training.
  • the target image recognition model obtained through pre-training can quickly create an image recognition engine, and use the image recognition engine to perform image recognition on the image of the target business.
  • the target concept label corresponding to the target service identifier may be determined by multi-user collaboration.
  • the image recognition method may further include the following steps: receiving a label collaboration request corresponding to the target service identifier; and obtaining the target concept label corresponding to the target service identifier according to the concept label selected on the collaboration terminal corresponding to the label collaboration request.
  • the tag collaboration request is used to request other users to collaborate to complete tag selection.
  • a cooperative terminal refers to a terminal that cooperates to complete label selection.
  • the target service terminal corresponding to the target service party can respond to the operation of the target service party to trigger the sending of a label collaboration request to the server.
  • the label collaboration request can carry the user identification of the collaboration user.
  • the server After the server receives the label collaboration request, it can send the identification of the collaboration user Joining the set of collaborative users corresponding to the target service ID, the server can send collaboration invitation information, such as an invitation link, to the collaboration terminal, and the collaboration terminal can enter the tab collaboration page in response to the collaboration user’s click operation on the invitation link to collaborate to complete the target service ID The selection of the corresponding concept label.
  • FIG. 9 it is a schematic diagram of an interface for requesting users to cooperate in some embodiments.
  • the collaboration configuration interface 904 displays the target service identification, collaboration user identification, and controls for adding collaboration users .
  • the shopping cart name "Canine Classifier” in Figure 9 represents the target business identifier.
  • the collaborating user's user IDs are displayed in the partner list: User 1 and User 2.
  • the "Add Partner" control is a control for adding collaborative users.
  • the collaborative terminal When the collaborative terminal receives a trigger operation on the control, such as the voice of "trigger adding a partner control", the collaborative terminal can display users who have an associated relationship with the target business party.
  • the identification for example, displays the identification of a user in the same company as the target business party, or the target business party's friend list in an instant messaging application.
  • the server may send the target service identification to the collaboration terminal corresponding to the collaboration user, and when the collaboration terminal receives the selection operation of the target service identification, it may send a request to the server to obtain the selected label set corresponding to the target service identification , The server can send the selected label set corresponding to the target service identifier to the collaboration terminal.
  • the selected label set includes the label corresponding to the target service identifier that has been selected by the user, which can be selected by the target user or selected by the collaborative user of.
  • the target service identifier and the corresponding selected label set are displayed in the collaboration terminal, which can make it easier for the collaboration user to understand what the target service is and the labels that the service has selected, thereby improving the efficiency of collaboration.
  • FIG. 10 a schematic diagram of an interface for a collaborative terminal to determine a collaborative task provided by some embodiments.
  • the collaboration terminal can display the area for selecting the target business logo, and display the prompt message "Please select your shopping cart".
  • "Created by friends” means that the shopping cart is a shopping cart created by other users and requires collaboration.
  • "Created by yourself” means a shopping cart created by the user himself.
  • the cooperative terminal receives a selection operation such as a click operation on the target service identifier "dog classifier”, it triggers to send a request to the server to obtain the selected label set corresponding to the "dog classifier", and the cooperative terminal receives the request sent by the server After the label set is selected, the labels in the selected label set are displayed on the interface. As shown in Figure 11. Since the shopping cart is created by a friend, the "training" and "preview” controls may not be displayed on the collaboration terminal.
  • the tag collaboration request carries the collaboration user ID, the target service ID, and the collaboration label range.
  • the image recognition method further includes: obtaining the label corresponding to the collaboration label range as the collaboration label; sending the collaboration label and the target service ID to the collaboration The collaboration terminal corresponding to the user ID, so that the collaboration terminal displays the target service ID and the corresponding collaboration label on the label collaboration selection interface.
  • the cooperation tag range is used to indicate the tag range during tag cooperation.
  • the scope of label collaboration can be indicated by the level information corresponding to the label.
  • the label collaboration scope can be the second level, which means that in the label tree, the label of the second level is within the scope of label collaboration. Collaborating users can select the second level.
  • Label The cooperative label range can be determined by the target business terminal in response to the target business party’s label range selection operation.
  • the target business party can determine the cooperative label range according to actual needs, and the target business terminal can display the level information of each label in the label tree.
  • the target service terminal determines the label cooperation range according to the selection operation of the hierarchical information.
  • Collaboration tags refer to tags that can be triggered by the collaboration terminal.
  • the non-cooperative tag is locked to the cooperative terminal.
  • the server may not send the non-cooperative tag to the cooperative terminal.
  • the collaboration label is sent to the collaboration terminal corresponding to the collaboration user ID, so that the collaboration terminal displays the collaboration label on the label collaboration selection interface, so that when the collaboration user performs label collaboration, If the concept tag is selected within the scope of collaboration, it can prevent the collaborative user from selecting invalid concept tags, and improve the efficiency of collaborative selection of tags.
  • the server may discard the non-cooperative label, that is, not add the non-cooperative label to the selected label set corresponding to the target service identifier.
  • obtaining the target concept label corresponding to the target service identifier according to the selected concept label triggered by the collaboration terminal corresponding to the label collaboration request includes: sending the concept label selected by the collaboration terminal corresponding to the label collaboration request to the target service terminal.
  • the concept label display area is used to display the selected concept label; when the target service terminal receives a trigger operation that triggers an image recognition service customization request, the concept label currently displayed in the concept label display area is used as the target service Identifies the corresponding target concept label.
  • the concept label display area displays the concept label corresponding to the target service identifier that has been selected by the user.
  • the server may synchronize the concept label selected by the collaboration terminal to the target service terminal, or may send it after receiving the label acquisition request sent by the target service terminal.
  • the target service terminal may trigger to send a label acquisition request to the server when it receives a trigger operation on the "refresh" control in FIG. 8.
  • the target service terminal When the target service terminal receives the selection operation of the "training" control, it triggers to send an image recognition service customization request to the server, and the target concept label carried in the image recognition service customization request is the concept label currently displayed in the concept label display area. For example, assuming that the concept label display area displays 5 concept labels, 4 of which are the image labels selected by the target business party and the other is the image label sent by the collaborative terminal, the 5 concept labels can be used as a "dog classifier" The corresponding target concept label.
  • the tags that have been added to the current shopping cart can be highlighted. For example, the tags that have been added to the shopping cart can be displayed in red in the tag tree, so that the business party can quickly determine that the tag has been added. Distribution.
  • the image recognition system mainly includes three modules: the design of the "label store", the automatic mining of labels and corresponding image data, and the automatic evolution of the image recognition basic model.
  • the image tag and the hierarchical information corresponding to the image tag may be preset by the service provider that provides the image recognition service.
  • the program developer of the service provider can build an image tag tree as needed.
  • the tag tree can be published in the tag store.
  • the server can automatically assign label numbers to the labels in the image label tree, and the server can also check whether the labels in the label tree are duplicated.
  • the server can also automatically obtain the image corresponding to the entity tag in the tag tree.
  • the first entity tag set includes entity tags corresponding to multiple concept tags; obtain the image corresponding to each first entity tag in the first entity tag set as a training image to obtain the training image gather.
  • the server may use the image tags of the leaf nodes in the image tag tree as the entity tags (first entity tags), that is, the lowest-level tags may be used as the entity tags to obtain the first entity tag set.
  • entity tags includes entity tags corresponding to a plurality of concept tags.
  • the first set of entity tags may include entity tags corresponding to birds and dogs, respectively.
  • the server can scan the service provider daily for adding new entity tags, and if so, update the tags. And obtain the original label and the newly mined image corresponding to the new label, and update the image database.
  • program developers can pre-design the process of offline image mining operations. This process can scan the tag tree once a day to find entity tags that do not have corresponding images, and then perform image data mining. For example, a large number of images and text corresponding to the image can be obtained, and the text corresponding to the image can be compared with the entity tag. If the text corresponding to the image includes an entity tag, the image is used as the image corresponding to the entity tag.
  • the training of the model may be training once every preset duration. For example, it can be training once a day.
  • Program developers can design an offline update process for the model. This process can automatically scan the tag tree at a certain point in time every day, pull all the entity tags with image data, and train based on these entity tags and the corresponding training image set. Generate a general image recognition model that can output the recognition results of each entity tag in the tag tree, replace the original general image recognition model, and complete the process of model replacement.
  • the server may determine whether the replacement image is a model according to the selection of the business party. For example, in order to ensure business stability, the business party may choose not to evolve. Then the server does not update the image recognition model in the image recognition engine corresponding to the business party.
  • the image recognition service customization request sent by the target service terminal is received, and the image recognition service customization request carries the target service identifier and the corresponding target concept label.
  • the target business party when it needs to customize the required image recognition service, for example, the image recognition service for identifying dogs, it can send an image recognition service customization request to the server by operating the target business terminal.
  • the image recognition service customization request carries the "canine classifier" and the corresponding target concept label. For example, shepherd dogs and poodles.
  • the target business party can add conceptual tags to the tag shopping cart by searching tags in the tag tree or importing tags to complete tag customization for a specific business.
  • the tags on the tag shopping cart can be done collaboratively by multiple people.
  • the shopping cart is divided into "my shopping cart” and "friend's shopping cart”.
  • Users can add or delete tags in "My Shopping Cart” to customize a specific image recognition engine, or delete their own shopping carts that have been created.
  • the user can also add tags to the "friend's shopping cart” for collaborative addition of tags, and can also pre-set the user to have the authority to add tags to the tags on the "friend's shopping cart” without the authority to delete tags.
  • "Friend's shopping cart” is designed to support simultaneous operation by multiple people, speed up the tag addition of shopping carts, and help shopping cart owners to establish a large number of target concept tags.
  • the image recognition platform can support the creation of image service engines corresponding to multiple business parties at the same time.
  • an image recognition engine can be created for the target business party, such as an image recognition service that recognizes dogs.
  • the basic image recognition model can output the image recognition results corresponding to each entity tag in the tag tree, so it can be based on the target concept tag in the shopping cart and the trained The basic image recognition model creates the image recognition service corresponding to the target business party.
  • the image recognition service may be provided in at least one form of a webpage, a parent application, or a sub-application.
  • sub-applications are commonly called applets, which are applications that can be implemented in the environment provided by the parent application.
  • the parent application is the application that carries the sub-application and provides an environment for the realization of the sub-application.
  • the parent application is a native application. Native applications are applications that can run directly on the operating system.
  • the parent application program may be an instant messaging application program
  • the child application program may be an image recognition program.
  • the icon of the child application can be displayed on the page corresponding to the parent application.
  • the child application can be run in the environment provided by the parent application, such as running an image recognition program , So that users can perform image recognition without having to install an image recognition application in the terminal in advance.
  • the image recognition request carrying the target service identifier is received, it is determined according to the corresponding relationship that the image recognition request is an image recognition request for the target concept label.
  • the image recognition result in the target image recognition result set is the image recognition result corresponding to the target entity tag, and the target entity tag The entity tag corresponding to the target concept tag.
  • the target image recognition model is entity tags corresponding to multiple concept tags
  • the corresponding training images are obtained through training. Therefore, the target image recognition model can output the probability corresponding to each entity tag.
  • the probabilities corresponding to "Olivia”, “Swallow”, “German Shepherd”, “Scottish Shepherd”, “Poodle”, “Poodle”, and “Teddy” can be output.
  • poodles, poodles, and teddy dogs are nicknames for poodles, that is, “poodles", “poodles”, and “teddy” are the physical tags corresponding to the concept label "poodles”.
  • the image recognition result of the target entity tag corresponding to the target concept tag "shepherd” includes the probabilities corresponding to "German Shepherd” and “Scottish Shepherd” respectively.
  • the image recognition result of the target entity tag corresponding to the target concept tag "Poodle” includes the probabilities corresponding to "Poodle”, “Poodle”, and “Teddy” respectively.
  • the largest probability among the probabilities corresponding to "German Shepherd” and “Scottish Shepherd” can be obtained as the probability corresponding to the target concept label "Shepherd Dog”.
  • the largest probability among the probabilities corresponding to "Poodle”, “Poodle”, and “Teddy” can be obtained as the probability corresponding to the target concept label "Poodle”. Therefore, even if the model that outputs the recognition probabilities of the concept labels of "Shepherd Dog” and “Poodle” is not trained, the probability of the concept labels of "Shepherd Dog” and “Poodle” can be obtained, which is equivalent to that the image recognition engine model can infer Obtain concept labels that have not been “seen” during the model training process.
  • the business side can also independently and quickly build the required image recognition engine services, which can save a lot of manpower and resources, so that the service provider can focus on developing image recognition models with higher recognition rates, and reduce response to various business needs.
  • the business side can focus on the business, and can also get the services that meet the business needs without understanding the specific details of image recognition.
  • an image recognition apparatus is provided.
  • the apparatus can adopt software modules or hardware modules, or a combination of the two can become a part of computer equipment.
  • the apparatus specifically includes: image recognition request The receiving module 1302, the target image acquisition module 1304, the target image recognition result set acquisition module 1306, and the result acquisition module 1308, wherein:
  • the image recognition request receiving module 1302 is configured to receive an image recognition request for the target concept tag.
  • the target image acquisition module 1304 is configured to acquire the target image to be recognized corresponding to the image recognition request.
  • the target image recognition result set acquisition module 1306 is used to input the target image into the target image recognition model corresponding to the entity tag for image recognition to obtain the target image recognition result set, and the image recognition result in the target image recognition result set is the target entity tag Corresponding to the image recognition result, the target entity label is the entity label corresponding to the target concept label.
  • the result obtaining module 1308 is configured to obtain the image recognition result corresponding to the target concept label according to the target image recognition result set.
  • the target image recognition result set obtaining module 1306 includes: a candidate image recognition result set obtaining unit, configured to input the target image into the target image recognition model corresponding to the entity tag for image recognition, and obtain the candidate image recognition result set ,
  • the candidate image recognition result set includes the image recognition result corresponding to the first entity tag;
  • the target entity tag determination unit is used to determine the first entity tag corresponding to the target concept tag as the target entity tag;
  • the target image recognition result set acquisition unit uses To obtain the image recognition result corresponding to the target entity tag from the candidate image recognition result set, to obtain the target image recognition result set.
  • the image recognition device further includes: a first entity tag set obtaining module, configured to obtain a first entity tag set, the first entity tag set includes entity tags corresponding to a plurality of concept tags; the training image set obtains The module is used to obtain the image corresponding to each first entity tag in the first entity tag set as a training image to obtain the training image set; the model training module is used to perform model training according to the training image set to obtain the target image recognition model,
  • the target image recognition model is regarded as a general image recognition model corresponding to multiple concept tags.
  • the image recognition result corresponding to the target entity tag includes the possibility that the target image includes the entity corresponding to the target entity tag
  • the result obtaining module 1308 is configured to: obtain from the target image recognition result set that satisfies the likelihood condition Possibility, as the target possibility; according to the target possibility, the image recognition result corresponding to the target concept label is obtained.
  • the conditions of the possibility include the possibility of sorting before the first sorting, the possibility of sorting as the first sorting, or the possibility of being greater than the first preset At least one of the thresholds.
  • the probability condition is the maximum probability
  • the result obtaining module 1308 is configured to perform at least one of the following steps: when the target probability is greater than the second preset threshold, determining the image recognition result corresponding to the target concept tag includes: The target image includes the entity corresponding to the target concept label; or the target possibility is used as the image recognition result corresponding to the target concept label.
  • the image recognition apparatus further includes: a service customization request receiving module, configured to receive an image recognition service customization request sent by a target service terminal, the image recognition service customization request carries the target service identifier and the corresponding target concept label; correspondence relationship The establishment module is used to establish the corresponding relationship between the target service identifier and the target concept label in response to the image recognition service customization request; the image recognition request receiving module is used to: when the image recognition request carrying the target service identifier is received, determine according to the corresponding relationship The image recognition request is an image recognition request for the target concept label.
  • a service customization request receiving module configured to receive an image recognition service customization request sent by a target service terminal, the image recognition service customization request carries the target service identifier and the corresponding target concept label
  • correspondence relationship The establishment module is used to establish the corresponding relationship between the target service identifier and the target concept label in response to the image recognition service customization request
  • the image recognition request receiving module is used to: when the image recognition request carrying the target service identifier is received, determine according to the corresponding relationship
  • the image recognition device further includes: a tag cooperation request receiving module, which is used to receive a tag cooperation request corresponding to the target service identifier; a target concept tag obtaining module, which is used to trigger selection on the cooperation terminal corresponding to the tag cooperation request To obtain the target concept label corresponding to the target business identifier.
  • the target concept label obtaining module is used to: send the concept label selected by the collaboration terminal corresponding to the label collaboration request to the concept label display area of the target service terminal, and the concept label display area is used to display the selected concept Label:
  • the target service terminal receives a trigger operation that triggers an image recognition service customization request
  • the concept label currently displayed in the concept label display area is used as the target concept label corresponding to the target service identifier.
  • the tag collaboration request carries the collaboration user ID, the target service ID, and the collaboration tag range.
  • the image recognition device further includes: a collaboration tag acquisition module for obtaining a tag corresponding to the collaboration tag range as a collaboration tag; sending collaboration information The module is used to send the collaboration label and the target service ID to the collaboration terminal corresponding to the collaboration user ID, so that the collaboration terminal displays the target service ID and the corresponding collaboration label on the label collaboration selection interface.
  • Each module in the above-mentioned image recognition device can be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the form of hardware or independent of the processor in the computer equipment, or may be stored in the memory of the computer equipment in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 14.
  • the computer equipment includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium.
  • the database of the computer equipment is used to store images.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions are executed by the processor to realize an image recognition method.
  • FIG. 14 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied.
  • the specific computer device may Including more or fewer parts than shown in the figure, or combining some parts, or having a different arrangement of parts.
  • a computer device including a memory and a processor, where computer-readable instructions are stored in the memory, and the processor implements the steps in the foregoing method embodiments when executing the computer-readable instructions.
  • a computer-readable storage medium is provided, and computer-readable instructions are stored, and when the computer-readable instructions are executed by a processor, the steps in the foregoing method embodiments are implemented.
  • a computer program product or computer program includes computer instructions stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the steps in the foregoing method embodiments.
  • Non-volatile memory may include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, or optical storage.
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)

Abstract

一种图像识别方法、装置、计算机设备和存储介质。涉及人工智能，所述方法包括：接收针对目标概念标签的图像识别请求（S202）；获取图像识别请求对应的待识别的目标图像（S204）；将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，目标实体标签为所述目标概念标签对应的实体标签（S206）；根据目标图像识别结果集合得到目标概念标签对应的图像识别结果（S208）。

Description

图像识别方法、装置、计算机设备和存储介质
本申请要求于2020年03月18日提交中国专利局，申请号为202010189599.3，申请名称为“图像识别方法、装置、计算机设备和存储介质”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
本申请涉及图像识别技术领域，特别是涉及一种图像识别方法、装置、计算机设备和存储介质。
背景技术
随着科技的发展，图像识别技术被越来越广泛的使用。图像识别技术是指识别出图像所包括的对象的技术，例如识别图像中是否包括牛。
相关技术中，可以通过基于人工智能的图像识别模型进行图像识别。图像识别模型需要预先根据训练数据进行模型训练。然而，不同的业务场景下所需要输出的图像标签经常存在差别，因此经常存在当需要进行图像识别，需要根据所需要识别的图像标签进行模型训练，再根据训练得到的图像识别模型进行图像识别的情况，浪费计算机资源，且导致图像识别效率低。
发明内容
根据本申请提供的各种实施例，提供一种图像识别方法、装置、计算机设备和存储介质。
一种图像识别方法，由计算机设备执行，所述方法包括：接收针对目标概念标签的图像识别请求；获取所述图像识别请求对应的待识别的目标图像；将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，所述目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，所述目标实体标签为所述目标概念标签对应的实体标签；根据所述目标图像识别结果集合得到所述目标概念标签对应的图像识别结果。
一种图像识别装置，所述装置包括：图像识别请求接收模块，用于接收针对目标概念标签的图像识别请求；目标图像获取模块，用于获取所述图像识别请求对应的待识别的目标图像；目标图像识别结果集合获取模块，用于将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，所述目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，所述目标实体标签为所述目标概念标签对应的实体标签；结果获取模块，用于根据所述目标图像识别结果集合得到所述目标概念标签对应的图像识别结果。
一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述计 算机可读指令被所述处理器执行时实现以下步骤：接收针对目标概念标签的图像识别请求；获取所述图像识别请求对应的待识别的目标图像；将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，所述目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，所述目标实体标签为所述目标概念标签对应的实体标签；根据所述目标图像识别结果集合得到所述目标概念标签对应的图像识别结果。
一个或多个存储有计算机可读指令的非易失性存储介质，所述计算机可读指令被一个或多个处理器执行时实现以下步骤：接收针对目标概念标签的图像识别请求；获取所述图像识别请求对应的待识别的目标图像；将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，所述目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，所述目标实体标签为所述目标概念标签对应的实体标签；根据所述目标图像识别结果集合得到所述目标概念标签对应的图像识别结果。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征、目的和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请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单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为一些实施例中图像识别方法的应用环境图；
图2为一些实施例中图像识别方法的流程示意图；
图3为一些实施例中图像标签树的示意图；
图4为另一些实施例中触发图像识别请求以及显示目标概念标签对应的图像识别结果的界面示意图；
图5为一些实施例中将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合的流程示意图；
图6为一些实施例中对图像识别模型进行模型训练，得到目标图像识别模型的流程图；
图7为一些实施例中标签修改界面的示意图；
图8为一些实施例中图像识别业务定制界面的示意图；
图9为一些实施例中请求用户进行协作的界面示意图；
图10为一些实施例中协作终端确定协作任务的界面示意图；
图11为一些实施例中协作界面的示意图；
图12为一些实施例中图像识别系统的运行原理图；
图13为一些实施例中图像识别装置的结构框图；以及
图14为一些实施例中计算机设备的内部结构图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
人工智能(Artificial Intelligence,AI)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。换句话说，人工智能是计算机科学的一个综合技术，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式做出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。
人工智能技术是一门综合学科，涉及领域广泛，既有硬件层面的技术也有软件层面的技术。人工智能基础技术一般包括如传感器、专用人工智能芯片、云计算、分布式存储、大数据处理技术、操作/交互系统、机电一体化等技术。人工智能软件技术主要包括计算机视觉技术、语音处理技术、自然语言处理技术以及机器学习/深度学习等几大方向。
计算机视觉技术(Computer Vision,CV)计算机视觉是一门研究如何使机器“看”的科学，更进一步的说，就是指用摄影机和电脑代替人眼对目标进行识别、跟踪和测量等机器视觉，并进一步做图形处理，使电脑处理成为更适合人眼观察或传送给仪器检测的图像。作为一个科学学科，计算机视觉研究相关的理论和技术，试图建立能够从图像或者多维数据中获取信息的人工智能系统。计算机视觉技术通常包括图像处理、图像识别、图像语义理解、图像检索、OCR、视频处理、视频语义理解、视频内容/行为识别、三维物体重建、3D技术、虚拟现实、增强现实、同步定位与地图构建等技术，还包括常见的人脸识别、指纹识别等生物特征识别技术。
机器学习(Machine Learning,ML)是一门多领域交叉学科，涉及概率论、统计学、逼近论、凸分析、算法复杂度理论等多门学科。专门研究计算机怎样模拟或实现人类的学习行为，以获取新的知识或技能，重新组织已有的知识结构使之不断改善自身的性能。机器学习是人工智能的核心，是使计算机具有智能的根本途径，其应用遍及人工智能的各个领域。机器学习和深度学习通常包括人工神经网络、置信网络、强化学习、迁移学习、归纳学习、式教学习等技术。
随着人工智能技术研究和进步，人工智能技术在多个领域展开研究和应用，例如常见的智能家居、智能穿戴设备、虚拟助理、智能音箱、智能营销、无人驾驶、自动驾驶、无人机、机器人、智能医疗、智能客服等，相信随着技术的发展，人工智能技术将在更多的领域得到应用，并发挥越来越重要的价值。
本申请实施例提供的方案涉及人工智能的图像识别等技术，具体通过如下实施例进行说明:
本申请提供的图像识别方法，可以应用于如图1所示的应用环境中。其中，终端102通过网络与服务器104进行通信。服务器104中预先部署有预先进行模型训练得到的目标图像识别模型。当需要进行图像识别时，用户可以通过操作终端102触发针对目标概念标签的图 像识别请求，终端102向服务器104发送针对目标概念标签的图像识别请求，服务器104执行本申请实施例提供的方法，得到目标概念标签对应的图像识别结果。服务器104将目标概念标签对应的图像识别结果返回至终端102，终端102显示目标概念标签对应的图像识别结果。可以理解，本申请实施例提供的图像识别方法还可以应用于终端102中。例如，目标图像识别模型可以部署于终端102中。
其中，终端102可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备，服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一些实施例中，如图2所示，提供了一种图像识别方法，以该方法应用于图1中的服务器为例进行说明，包括以下步骤：
步骤S202，接收针对目标概念标签的图像识别请求。
其中，概念标签用于表示概念，实体标签用于表示实体。概念是对具有共同的特点的实体的表达。实体是指概念对应的具体的事物。概念包括它所定义的各个实体，一个概念标签下对应多个实体标签，“多个”是指“至少两个”。例如，概念标签可以是汽车，实体标签则可以包括A品牌的汽车以及B品牌的汽车。又例如，概念标签可以是犬类，实体标签可以包括牧羊犬以及猎犬等。
可以理解，概念标签以及实体标签是相对的，一个图像标签是概念标签还是实体标签，可以根据实际情况变化例如应用场景变化。例如，对于两个概念，如果前一个概念包括后一个概念所表达的意思，则前一个概念相对于后一个概念，为概念标签。后一个概念相对于前一个概念，为实体标签。例如，在C场景中，“汽车”标签是实体标签。“交通工具”标签是概念标签。在D场景中，“汽车”标签是概念标签，汽车的品牌例如“A品牌”以及“B品牌”是实体标签。
图像识别请求用于请求对图像的内容进行识别。针对目标概念标签的图像识别请求是指：该图像识别请求用于请求识别图像中目标概念标签对应的内容。例如，针对“汽车”标签的图像识别请求用于请求识别图像中是否包括汽车。
具体地，目标概念标签可以为一个或多个，图像识别请求中可以携带目标概念标签对应的标签相关信息，该标签相关信息可以包括概念标签标识或者目标业务标识。当服务器接收到携带概念标签标识的图像识别请求时，获取概念标签标识对应的标签，作为目标概念标签，确定该图像识别请求为针对目标概念标签的图像识别请求。也可以预先建立业务标识与概念标签的对应关系，当接收到携带目标业务标识的图像识别请求时，服务器可以根据该对应关系确定目标概念标签，确定图像识别请求为针对目标概念标签的图像识别请求。
在一些实施例中，服务器中可以预先存储图像标签之间的层级关系，可以根据图像标签的层级确定图像标签的类型，一个图像标签的类型可以为概念标签或者实体标签。例如可以将层级最低的图像标签作为实体标签，将实体标签对应的上一层级的图像标签作为概念标签。
在一些实施例中，可以将对图像识别模型进行模型训练时，用于模型训练的训练图像对应的图像标签作为实体标签，表示实体对应的上位概念的标签为概念标签。训练图像对应的 图像标签是指该训练图像中包括该图像标签直接对应的实体。
例如，如图3所示，为一些实施例中图像标签树的示意图。图像标签按照层级关系可以以树形结构展示，包括非叶子节点和叶子节点。叶子节点是指没有子节点的节点，即层级最低的节点，例如图3中的C1、C2、C3以及B2节点。非叶子节点的标签表示某种或粗或细的概念，因此被称为“概念标签”，叶子节点的标签表示特定的对应有训练图像的标签，因此被称为“实体标签”，在训练模型时，可以获取实体标签对应的训练图像，进行模型的训练。
步骤S204，获取图像识别请求对应的待识别的目标图像。
具体地，图像识别请求中可以携带图像或者图像索引信息。当图像识别请求携带图像时，将该图像作为待识别的目标图像。图像索引信息例如可以是图像的存储位置或者图像的标识例如名称。当该图像识别请求携带图像索引信息时，可以根据图像索引信息获取得到对应的图像，作为待识别的目标图像。
举个实际的例子，假设“acc”文件夹存储有5张图片，当用户希望对该5张图片进行图像识别时，则图像识别请求中可以携带文件夹的名称“acc”，服务器可以将“acc”文件夹里的5张图片作为待识别的目标图像。
步骤S206，将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，目标实体标签为目标概念标签对应的实体标签。
其中，图像识别模型用于对图像进行识别，图像识别模型例如可以是卷积神经网络模型。实体标签对应的目标图像识别模型是指：该目标图像识别模型输出的是实体标签对应的识别结果。目标实体标签是目标概念标签对应的实体标签，目标实体标签所对应的实体，是属于目标概念的实体。例如，目标概念标签为“计算机设备”，则目标实体标签可以为“平板电脑”以及“服务器”。目标实体标签对应的图像识别结果表示目标图像中包括目标实体标签对应的实体的情况。目标实体标签对应的图像识别结果可以为模型输出的目标图像中包括实体标签对应的实体的可能度，其中可能度是指图像中包括实体的可能程度，实体标签对应的可能度越大，则表示图像中越可能存在实体标签对应的实体，可能度可以通过概率表示。
目标图像识别模型可以是预先通过机器学习训练得到的，可以采用有监督的训练方法，获取训练图像以及对应的实体标签，进行模型训练。模型训练的目标可以是使模型输出的实体标签对应的概率分布与训练图像对应的实体标签对应的概率分布的差异越小越好，即可以朝着使模型损失值下降的方向调整模型的参数，直至模型收敛，得到目标图像识别模型。模型输出的实体标签对应的概率分布与训练图像对应的实体标签对应的概率分布的差异，与模型损失值可以成正相关关系，即差异越大，则模型损失值越大。
具体地，得到目标图像后，服务器可以将目标图像输入到目标图像识别模型中，目标图像识别模型可以对图像进行处理，例如经过卷积层、全连接层以及激活层的处理，输出各个实体标签对应的概率，可以获取目标实体标签对应的概率，组成目标图像识别结果集合。
步骤S208，根据目标图像识别结果集合得到目标概念标签对应的图像识别结果。
其中，目标概念标签对应的图像识别结果可以是目标概念标签对应的可能度以及目标概念标签对应的实体的判断结果的至少一个。例如，当目标概念标签为“交通工具”时，目标概念标签对应的图像识别结果可以是“图像中包括交通工具的概率为0.9”，也可以是判断结果“图像中包括交通工具”。
具体地，得到目标图像识别结果集合后，由于该目标图像识别结果集合中的图像识别结果为目标实体标签对应的识别结果，目标概念标签所表示的概念包括目标实体标签所表示的实体，因此服务器可以根据目标图像识别结果集合得到目标概念标签对应的图像识别结果。
举个实际的例子，假设目标概念标签为“交通工具”，目标图像识别模型为识别图像中是否包括“汽车”、“飞机”或者“轮船”等实体的识别模型。将目标图像输入到目标图像识别模型中，输出“汽车”、“飞机”、“轮船”分别对应的概率为“0.9”、“0.05”、“0.05”。则目标实体标签对应的图像识别结果可以为“0.9”、“0.05”、“0.05”。由于根据“0.9”、“0.05”、“0.05”得到最大概率为“汽车”对应的概率“0.9”，大于预设概率例如0.8，则目标概念标签对应的图像识别结果可以为目标图像中包括“交通工具”。
上述图像识别方法，对于针对目标概念标签的图像识别请求，对于针对目标概念标签的图像识别请求，可以将待识别的图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标实体标签对应的图像识别结果，根据目标实体标签对应的图像识别结果得到目标概念标签对应的图像识别结果。因此，当需要针对目标概念标签进行图像识别时，无需额外针对目标概念标签进行模型的训练，也能够得到目标概念标签对应的图像识别结果，减少了对计算机资源的占用，提高了图像识别效率。
在一些实施例中，目标实体标签对应的图像识别结果包括：目标图像中包括目标实体标签对应的实体的可能度，根据目标图像识别结果集合得到概念标签对应的图像识别结果包括：从目标图像识别结果集合中获取满足可能度条件的可能度，作为目标可能度；根据目标可能度得到目标概念标签对应的图像识别结果，可能度条件包括可能度排序在第一排序之前、可能度排序为第一排序或者可能度大于第一预设阈值的至少一个。
具体地，第一排序以及第一预设阈值可以根据需要设置，例如可能度条件可以为可能度最大，也可以是可能度大于0.8。得到目标可能度之后，可以根据目标可能度计算得到目标概念标签对应的可能度，也可以是根据目标可能度确定目标概念标签对应的实体的判断结果。例如，假设可能度条件包括可能度大于第一预设阈值，如果大于第一预设阈值的可能度大于预设数量，例如3，则说明该图像中存在目标概念标签对应的实体的可能性是比较大的，则可以确定目标概念标签对应的图像识别结果为：目标图像中包括目标概念标签对应的实体。
在一些实施例中，可能度条件可以为可能度最大，根据目标可能度得到目标概念标签对应的图像识别结果包括以下步骤的至少一个：当目标可能度大于第二预设阈值时，确定目标概念标签对应的图像识别结果包括目标图像中包括目标概念标签对应的实体；或者将目标可能度作为目标概念标签对应的图像识别结果。
具体地，第二预设阈值也可以根据需要设置。例如，可以是0.85。获取得到最大可能度后，可以将该最大可能度作为目标概念标签对应的图像识别结果，也可以是判断目标可能度 大于第二预设阈值时，确定目标图像中包括目标概念标签对应的实体。举个实际的例子，假设目标概念标签为“交通工具”，目标图像识别结果集合包括输出“汽车”、“飞机”、“轮船”分别对应的概率为“0.9”、“0.05”、“0.05”。则可以将最大概率“0.9”，作为目标图像中包括“交通工具”的概率。也可以判断最大概率“0.9”大于第二预设概率“0.85”时，确定目标图像中包括交通工具。
在一些实施例中，服务器得到目标概念标签对应的图像识别结果后，可以将该图像结果返回至发送图像识别请求对应的目标业务终端。例如，如图4所示，为一些实施例中触发图像识别请求以及显示目标概念标签对应的图像识别结果的界面示意图。该界面包括图片上传区域402、判断结果显示区域404以及概率显示区域406。“XX业务分类器”为图像识别业务的业务名称。当用户需要对图像进行识别时，可以点击“上传”按钮，进入图像上传界面选择图片，当图片选择完毕，接收到确认操作后，目标业务终端可以触发向服务器发送针对目标概念标签的图像识别请求，目标概念标签例如可以为“鸟类”以及“鱼类”。服务器中部署有输出“鸟类”对应的实体标签例如“麻雀”、“雁”、“鹧鸪”等，以及“鱼类”对应的实体标签例如“金鱼”、“鲤鱼”以及“草鱼”等图像标签的概率的图像识别模型，服务器将终端上传的图片分别输入到图像识别模型中，该图像识别模型输出上传的图片中包括“麻雀”、“雁”、“鹧鸪”、“金鱼”、“鲤鱼”以及“草鱼”等的概率。服务器可以获取实体标签“麻雀”、“雁”、“鹧鸪”分别对应的概率中的最大概率，作为上传的图片中包括目标概念标签“鸟类”的概率。服务器可以获取实体标签“金鱼”、“鲤鱼”以及“草鱼”分别对应的概率中的最大概率，作为上传的图片中包括目标概念标签“鱼类”的概率。如果“鸟类”的概率为0.9667，大于预设概率0.85，服务器还可以输出“图片中包括鸟类”的图像识别结果。
在一些实施例中，如图5所示，步骤S206即将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合包括以下步骤：
步骤S502，将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到候选图像识别结果集合，候选图像识别结果集合包括各个第一实体标签对应的图像识别结果。
其中，第一实体标签可以是任意的实体标签。目标图像识别模型可以是通用的图像识别模型，可以用于输出图像中包括多个概念标签分别对应的实体标签的概率。例如，图像识别业务平台中，不同的业务方的图像识别需求是不同的，有些业务方的业务与游戏相关，因此需要识别的图像标签与游戏相关。有些业务方的业务与植物相关，需要识别的图像标签与植物相关。则可以部署一个既可以识别与植物相关的图像标签的模型，也可以识别游戏相关的图像标签的通用图像识别模型。
具体地，可以将目标图像输入到目标图像识别模型中，目标图像识别模型可以对图像进行处理，输出各个第一实体标签对应的概率，服务器可以获取第一实体标签对应的概率，组成候选图像识别结果集合。
步骤S504，确定目标概念标签对应的第一实体标签，作为目标实体标签。
具体地，可以预先设置概念标签与实体标签的对应关系，得到目标概念标签后，获取目 标概念标签对应的实体标签，作为目标实体标签。例如目标概念标签为“计算机设备”，第一实体标签包括“麻雀”、“雁”、“鹧鸪”、“金鱼”、“鲤鱼”、“草鱼”、“平板电脑”以及“服务器”，则目标概念标签“计算机设备”对应的目标实体标签可以为“平板电脑”以及“服务器”。
步骤S506，从候选图像识别结果集合中获取目标实体标签对应的图像识别结果，得到目标图像识别结果集合。
具体地，服务器得到目标实体标签后，可以从候选图像识别结果集合中，获取目标实体标签对应的图像识别结果，组成目标图像识别结果集合。
本申请实施例中，通过将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到各个第一实体标签对应的图像识别结果组成的候选图像识别结果集合，从候选图像识别结果集合中获取目标实体标签对应的图像识别结果，因此针对任意的目标概念标签，可以输入到相同的图像识别模型即通用识别模型中进行识别，从而无需为每个概念标签对应的实体标签训练一个专门的图像识别模型，节省了模型训练资源。
在一些实施例中，如图6所示，为一些实施例中对图像识别模型进行模型训练，得到目标图像识别模型的流程图，包括以下步骤：
步骤S602，获取第一实体标签集合，第一实体标签集合中包括多个概念标签分别对应的实体标签。
具体地，第一实体标签集合可以是预先设置的，可以是由具有图像识别需求的业务方人工设置的，也可以是由提供图像识别服务的服务方设置的。例如在图像识别平台中，服务方和业务方可在该平台以各自的身份登陆进行操作。服务方可以通过整理标签，将标签发布于“标签商店”中。业务方也可以根据需要添加标签。可以理解，该多个概念标签中包括目标概念标签。
在一些实施例中，服务方在整理以及发布标签过程中可以多人协作，各领域专家负责各自领域的标签，既减轻工作量，又确保专业性。添加或者修改标签的服务方作者名字可以在光标悬浮窗中出现，供其他的协作者查看，确保责任明晰。标签的数量可以根据实际情况设置，例如可以达到数万乃至数十万，标签可以树形结构展示层级关系，层级关系有助于理清体系脉络、快速定位标签，方便整理和发布标签。
在一些实施例中，服务方整理以及发布的标签可以包括概念标签以及实体标签，服务器可以接收修改请求，对概念标签以及实体标签进行修改。修改的类型可以包括增加标签、删除标签或者改变标签信息的至少一种。增加标签是指增加概念标签或者实体标签到指定的父节点下。服务器可以根据用户的删除操作删除标签，也可以是限制标签的删除，例如设预先置根节点的图像标签不可以删除。在当接收到删除操作，终端可以弹出询问框，询问是否删除标签，当接收到确认操作后，再进行删除避免误操作。改变标签信息可以是修改图像标签的描述，也可以是修改该标签的父亲节点。
如图7所示，为一些实施例中标签修改界面的示意图。当接收到针对标签“C2”的预设操作例如右键点击操作后，可以弹出标签修改界面，界面中显示“添加子节点”、“更新节 点”、“删除节点”以及“复制本标签编码”功能控件，通过触发对应的控件，可以实现对应的功能。
在一些实施例中，可以通过颜色表示标签节点的类型，不同颜色的标签节点具有不同的类型。例如普通节点为灰色，表示该标签为新增的标签，没有获取得到对应的图像。黄色节点表示支持查看该标签的图像数据，蓝色节点支持查看图像，还表示该标签为实体标签，已经根据实体标签对应的图像进行了模型的训练，故图像识别模型可以支持输出实体标签的对应的概率。对于支持查看图像的标签，可以在接收到对该标签的点击操作时，在终端上显示图像显示界面，该界面上可以显示图像以及图像的标注结果。
步骤S604，获取第一实体标签集合中各个第一实体标签分别对应的图像，作为训练图像，得到训练图像集合。
具体地，第一实体标签对应的图像指该图像中包括第一实体标签所表示的实体。例如，第一实体标签包括“麻雀”以及“手机”。则可以获取包括麻雀的图像以及获取包括手机的图像，将这些图像作为训练图像，得到训练图像集合。
步骤S606，根据训练图像集合进行模型训练，得到目标图像识别模型，将目标图像识别模型作为多个概念标签对应的通用图像识别模型。
其中，将目标图像识别模型作为多个概念标签对应的通用图像识别模型是指：目标图像识别模型是通用的，对于针对该多个概念标签中任一个概念标签的图像识别请求，可以获取对应的目标图像，输入到通用图像识别模型中进行图像识别。例如，假设多个概念标签包括“鸟类”标签以及“犬类”标签，则接收到针对“鸟类”标签的图像识别请求后，将对应的目标图像输入到该通用图像识别模型中进行图像识别。接收到针对“犬类”标签的图像识别请求后，可以将对应的目标图像输入到该通用图像识别模型中进行图像识别。
具体地，服务器得到训练图像集合后，可以将训练图像集合中的训练图像输入到待训练的图像识别模型中，模型输出该训练图像包括各个实体标签对应的实体的概率，可以根据模型输出的概率分布，以及训练图像对应的真实的图像标签的概率分布的差异，得到模型损失值，采用随机梯度下降方法调整模型的参数，直至模型收敛，得到目标图像识别模型，由于目标图像识别模型是根据多个概念标签分别对应的实体标签对应的图像训练得到的，因此可以将该目标图像识别模型作为多个概念标签对应的通用图像识别模型，当接收到针对这些概念标签的图像识别请求时，可以输入到该通用图像识别模型中进行图像的识别。
本申请实施例中，通过获取多个概念标签分别对应的实体标签＝所对应的图像进行模型的训练，可以训练得到对该多个概念标签的实体标签所表示的实体进行识别的图像识别模型，将该图像识别模型作为多个概念标签对应的通用图像识别模型，对于具有不同图像识别需求的业务方，也可以使用相同的模型进行图像的识别，得到实体标签的识别结果，再根据实体标签的识别结果得到目标概念标签对应的图像识别结果，因此可以在无需为不同的业务方专门训练对应的模型的情况下，满足不同的业务方的图像识别业务需求。
例如，假设A业务方的图像识别业务需求为鸟类图像识别业务，B业务方对应的图像识别业务需求为犬类图像识别业务。则可以训练得到一个既能识别得到鸟类图像识别业务对应 的实体标签以及犬类图像识别业务对应的实体标签对应的概率的通用图像识别模型，这样，对于A业务方以及B业务方的终端分别发送的图像识别请求，都可以获取图像识别请求对应的目标图像，输入到该通用图像识别模型中进行图像识别。
在一些实施例中，第一实体标签分别对应的图像可以是预先存储的，也可以是通过图像挖掘得到的，可以是进行离线的图像挖掘。例如，可以获取候选图像集以及图像对应的文本，候选图像集可以是在互联网上爬取得到的图像。可以每隔预设时长例如24小时扫描图像标签的标签树，查找其中的所有实体标签或者不存在对应的图像的实体标签，获取图像对应的文本中存在该实体标签的图像，作为该实体标签对应的图像，得到训练图像集合。例如，假设获取到图像t1，图像上的文本或者图像对应的介绍文本为“平板电脑”，则该图像t1为实体标签“平板电脑”对应的图像。这样，服务器可以自动获取训练图像集，进行模型的训练，使得目标图像识别模型可以随图像数据的累积进行潜在的进化，可以利用进化得到的目标图像识别模型更新图像识别引擎中旧的目标图像识别模型，以为业务方提供更好的服务。其中，图像识别引擎是指将图像作为输入，经过模型推断，输出图像标签的核心组件服务。
在一些实施例中，当挖掘得到第一实体标签分别对应的图像后，服务器可以发送标签颜色改变指令，以改变标签树中标签的颜色。例如可以使得灰色的叶子节点(无对应的图像的实体标签)变成黄色(有对应的图像的实体标签)。，使得业务方可以根据标签的颜色确定该标签是否存在对应的图像数据。服务器还可以根据标签颜色改变指令激活实体标签对应的图像查看功能，例如实体标签对应的“查看图像”功能控件，在接收到标签颜色改变指令之后，可以由灰色(表示不可点击)变更为黑色(表示可点击)。
在一些实施例中，图像识别方法还可以包括：接收目标业务终端发送的图像识别业务定制请求，图像识别业务定制请求携带目标业务标识以及目标概念标签；建立目标业务标识与目标概念标签的对应关系。因此当接收到携带目标业务标识的图像识别请求时，可以根据对应关系确定图像识别请求为针对目标概念标签的图像识别请求。
具体地，业务标识可以用于标识对应的业务，业务标识可以根据需要设置，可以是服务器自动生成的，也可以是根据业务方的业务标识输入操作得到的。例如，当目标业务方需要定制识别图像中是否包括“交通工具”的图像识别业务，则可以输入“交通工具分类器”，当目标业务方登录的目标业务终端接收到目标业务方触发图像定制业务的操作时，触发图像识别业务定制请求，并携带目标业务标识“交通工具分类器”。该图像识别业务定制请求用于请求定制“交通工具分类器”对应的图像识别业务。当接收到图像识别业务定制请求，服务器可以将目标业务标识与目标概念标签关联存储，这样，当接收到携带目标业务标识的图像识别请求时，服务器根据对应关系确定图像识别请求为针对目标概念标签的图像识别请求。因此，用户无需在每次需要进行图像识别时选择目标概念标签，提高了图像识别的效率。
例如，目标业务终端上可以显示业务定制入口“定制我的图像识别业务”，当接收到对业务定制入口的触发操作例如点击操作，可以进入图像识别业务定制界面，如图8所示，图像识别业务定制界面可以包括图像标签树显示区域802、概念标签显示区域804以及业务标识显示区域806。图像标签树显示区域802用于显示候选的图像标签，该候选的图像标签以 标签树的形式显示，目标业务终端可以接收用户的标签选择操作例如点击标签的操作。其中，可以设置目标业务方可以选择的标签为非叶子节点的标签，即概念标签。终端将用户选择的标签加入到概念标签显示区域804中进行显示，业务标识显示区域806显示用户当前选择的业务标识，例如“犬类分类器”为业务标识，表示该业务为对犬类进行分类的业务。如果用户需要再创建另一个业务对应的分类器，可以点击创建新业务的控件“创建新的购物车”，当目标业务终端接收到对“创建新的购物车”控件的选择操作时，目标业务终端可以显示业务标识输入框，目标业务终端获取用户在业务标识输入框输入的业务标识，作为新的业务标识。其中，“购物车”表示用户选择的概念标签的集合。当目标业务方对应的目标业务终端接收到对“训练”控件的触发操作，例如点击操作或者语音操作时，可以向服务器发送图像识别业务定制请求，服务器生成该图像识别业务定制请求对应的图像识别引擎。
此外，目标业务终端还可以接收对“刷新”控件的选择操作，服务器根据对“刷新”控件的选择操作对该业务标识对应的已选择的标签进行刷新，目标业务终端可以接收对“预览”控件的选择操作，向服务器发送图像识别引擎测试请求，服务器可以对生成的图像识别引擎进行测试。由于目标图像识别模型已经预先训练完毕，则接收到对“训练”控件的点击操作后，目标业务终端触发图像识别业务定制请求，服务器可以直接获取已训练得到的目标图像识别模型，作为该目标业务标识对应的图像识别引擎中的图像识别引擎，这样，可以提高构建图像识别引擎的效率。
其中，选择操作是用于选择目标信息的操作，可以是通过控件、语音、手势或者表情中的一种或多种操作选择信息。
在一些实施例中，服务器还可以接收目标业务方对应的目标业务终端发送的标签查询请求，查询请求可以携带查询信息，查询信息例如可以是标签的名称信息或者编码，服务器可以根据查询信息将对应的标签返回至目标业务终端，使得目标业务方可以快速进行标签的查询。例如，图像识别平台可以提供精确查找和模糊查找图像标签的功能，以供有明确标签列表的业务方快速查询标签，加快标签购物车的添加。其中，目标业务方可以通过图像标签名称，以及图像标签编码进行标签的查询。
在一些实施例中，目标图像识别模型可以是在接收到图像识别业务定制请求后，再进行模型训练得到的。目标图像识别模型也可以是预先已经训练得到的，通过预先训练得到目标图像识别模型，可以快速创建图像识别引擎，利用该图像识别引擎对目标业务的图像进行图像识别。
在一些实施例中，为了提高确定目标业务标识对应的目标概念标签的效率，目标业务标识对应的目标概念标签可以是多用户协作确定的。图像识别方法还可以包括以下步骤：接收目标业务标识对应的标签协作请求；根据在标签协作请求对应的协作终端上触发选择的概念标签，得到目标业务标识对应的目标概念标签。
具体地，标签协作请求用于请求其他用户协作完成标签的选择。协作终端是指协作完成标签选择的终端。目标业务方对应的目标业务终端可以响应于目标业务方的操作，触发向服务器发送标签协作请求，标签协作请求中可以携带协作用户的用户标识，服务器接收标签协 作请求后，可以将协作用户的标识加入目标业务标识对应的协作用户集合中，服务器可以向协作终端发送协作邀请信息，例如邀请链接，协作终端可以响应于协作用户对邀请链接的点击操作，进入标签协作页面，以协作完成目标业务标识对应的概念标签的选择。
如图9所示，为一些实施例中请求用户进行协作的界面示意图。当目标业务方对应的目标业务终端接收到对协作请求控件902的触发操作，例如点击操作时，显示协作配置界面904，协作配置界面904上显示目标业务标识、协作用户标识以及添加协作用户的控件。例如，图9中购物车名称“犬类分类器”表示目标业务标识。合作者列表中显示协作用户的用户标识：用户1和用户2。“添加合作者”控件为添加协作用户的控件，当协作终端接收到对该控件的触发操作，例如“触发添加合作者控件”的语音时，协作终端可以显示与目标业务方具有关联关系的用户标识，例如显示与目标业务方同一个公司的用户的标识，或者目标业务方在即时通讯应用中的好友列表。
在一些实施例中，服务器可以向协作用户对应的协作终端发送目标业务标识，当协作终端接收到对目标业务标识的选择操作时，可以向服务器发送获取目标业务标识对应的已选择标签集合的请求，服务器可以将目标业务标识对应的已选择标签集合发送到协作终端中，已选择标签集合中包括目标业务标识对应的已被用户选择的标签，可以是目标用户选择的，也可以是协作用户选择的。协作终端中显示目标业务标识以及对应的已选择标签集合，可以使得协作用户方便了解目标业务是什么以及该业务已经选择的标签，提高了协作的效率。
如图10所示，为一些实施例提供的协作终端确定协作任务的界面示意图。协作终端上可以显示选择目标业务标识的区域，并显示提示信息“请选择您的购物车”。“朋友创建”表示该购物车是其他用户创建的购物车，需要进行协作。“自己创建”表示是用户本人创建的购物车。当协作终端接收到对目标业务标识“犬类分类器”的选择操作例如点击操作时，触发向服务器发送获取“犬类分类器”对应的已选择标签集合的请求，协作终端接收到服务器发送的已选择标签集合后，在界面上显示已选择标签集合中的标签。如图11所示。由于该购物车是朋友创建的，因此协作终端上可以不显示“训练”以及“预览”控件。
在一些实施例中，标签协作请求携带协作用户标识、目标业务标识以及协作标签范围，图像识别方法还包括：获取协作标签范围对应的标签，作为协作标签；将协作标签以及目标业务标识发送至协作用户标识对应的协作终端，以使得协作终端在标签协作选择界面上显示目标业务标识以及对应的协作标签。
具体地，协作标签范围用于指示进行标签协作时，标签的范围。标签协作的范围可以是通过标签对应的层级信息表示的，例如，标签协作范围可以为第二层级，表示标签树中，第二层级的标签在标签协作范围内，协作用户可以选择第二层级的标签。协作标签范围可以是目标业务终端响应于目标业务方的标签范围选择操作确定的，例如，目标业务方可以根据实际需要确定协作标签范围，目标业务终端上可以显示标签树中各个标签的层级信息，目标业务终端根据对层级信息的选择操作确定标签协作范围。协作标签是指协作终端可以触发选择的标签。非协作标签对于协作终端而言，是处于锁定状态的，例如服务器可以不将非协作标签发送至协作终端中。通过获取协作标签范围对应的标签，作为协作标签，将协作标签发送 至协作用户标识对应的协作终端，以使得协作终端在标签协作选择界面上显示协作标签，使得协作用户在进行标签协作时，是在协作范围内选择概念标签的，可以避免协作用户选择无效的概念标签，提高了协作选择标签的效率。
在一些实施例中，对于协作终端发送的非协作标签，服务器可以将该非协作标签丢弃，即不加入到目标业务标识对应的已选择标签集合中。
在一些实施例中，根据标签协作请求对应的协作终端触发选择的概念标签，得到目标业务标识对应的目标概念标签包括：将标签协作请求对应的协作终端触发选择的概念标签，发送至目标业务终端的概念标签显示区域中，概念标签显示区域用于显示已选择的概念标签；将目标业务终端接收到触发图像识别业务定制请求的触发操作时，概念标签显示区域当前显示的概念标签，作为目标业务标识对应的目标概念标签。
具体地，概念标签显示区域显示的是目标业务标识对应的已被用户选择的概念标签。服务器可以是将协作终端触发选择的概念标签同步至目标业务终端中，也可以是接收到目标业务终端发送的标签获取请求时，再进行发送。例如目标业务终端可以是当接收到对图8中“刷新”控件的触发操作时，触发向服务器发送标签获取请求。
当目标业务终端接收到对“训练”控件的选择操作时，触发向服务器发送图像识别业务定制请求，图像识别业务定制请求携带的目标概念标签为概念标签显示区域当前显示的概念标签。例如，假设概念标签显示区域显示5个概念标签，其中4个为目标业务方选择的图像标签，另一个为协作终端发送的图像标签，则可以将该5个概念标签作为“犬类分类器”对应的目标概念标签。
在一些实施例中，对于已经加入当前购物车的标签，可以进行突出显示，例如对于已被加入到购物车的标签，在标签树中可以显示为红色，以使得业务方可以快速确定已加入标签的分布情况。
以下以创建犬类分类业务为例，结合图12所示的图像识别系统的运行原理图，对本申请实施例提供的图像识别方法进行说明。如图12所示，图像识别系统主要包括“标签商店”的设计、标签以及对应的图像数据的自动挖掘和图像识别基础模型自动进化这三个模块。
1、获取图像标签以及图像标签对应的层级信息。
具体地，图像标签以及图像标签对应的层级信息可以是提供图像识别服务的服务提供方预先设置的。例如，服务提供方的程序开发人员可以根据需要建立图像标签树。该标签树可以发布于标签商店中。服务器可以为图像标签树中的标签自动分配标签的编号，服务器也可以对标签树中的标签是否重复进行检查。服务器还可以自动获取标签树中实体标签对应的图像。
2、获取第一实体标签集合，第一实体标签集合中包括多个概念标签分别对应的实体标签；获取第一实体标签集合中各个第一实体标签分别对应的图像，作为训练图像，得到训练图像集合。
具体地，服务器可以将图像标签树中，叶子节点的图像标签作为实体标签(第一实体标签)，即可以将最低层级的标签作为实体标签，得到第一实体标签集合。该第一实体标签集 合包括多个概念标签对应的实体标签。例如，第一实体标签集合可以包括鸟类以及犬类分别对应的实体标签。
如图12所示，服务器可以每日扫描服务提供方是否有添加新的实体标签，如果有，则更新标签。并获取原有的标签以及新的标签对应的新挖掘的图像，进行图像库的更新。
例如，程序开发人员可以预先设计离线的图像自动挖掘作业的流程。该流程可以每日扫描一次标签树，查找其中的不存在对应的图像的实体标签，进行图像数据的挖掘。例如可以获取海量的图像以及图像对应的文本，将图像对应的文本与实体标签进行对比，如果图像对应的文本包括实体标签，则将该图像作为该实体标签对应的图像。
3、根据训练图像集合进行模型训练，得到目标图像识别模型，将目标图像识别模型作为多个概念标签对应的通用图像识别模型。
具体地，模型的训练可以是每隔预设时长训练一次。例如，可以是每天训练一次。程序开发人员可以设计模型离线更新流程，该流程定期例如在每日的某个时间点可以自动扫描标签树，拉取具有图像数据的所有实体标签，根据这些实体标签和对应的训练图像集合，训练生成可以输出标签树中各个实体标签的识别结果的通用图像识别模型，替代原有的通用图像识别模型，完成模型更替的过程。可以理解，服务器可以根据业务方的选择确定是否更替图像是被模型，例如，为了保证业务稳定，业务方可以选择不进化。则服务器不对该业务方对应的图像识别引擎中的图像识别模型进行更新。
4、接收目标业务终端发送的图像识别业务定制请求，图像识别业务定制请求携带目标业务标识以及对应的目标概念标签。
具体地，当目标业务方需要定制所需要的图像识别服务，例如识别犬类的图像识别服务时，可以通过操作目标业务终端，向服务器发送图像识别业务定制请求。图像识别业务定制请求携带“犬类分类器”以及对应的目标概念标签。例如，牧羊犬以及贵宾犬等。
例如，目标业务方可以通过搜索标签树的标签或者导入标签的方式，将概念标签加入到标签购物车中，完成特定业务的标签定制。标签购物车上的标签可以是多人协作完成的。例如，购物车分为“我的购物车”和“朋友的购物车”。用户本人可以在“我的购物车”可以进行标签的增删，以定制特定的图像识别引擎，也可以删除已经创建的自己的购物车。用户也可以在“朋友的购物车”上增加标签，以进行标签的协作添加，还可以预先设置用户对与“朋友的购物车”上的标签有增加标签的权限，无删除标签的权限，以更好的管理标签。“朋友的购物车”的设计主旨是为了支持多人同时操作，加快购物车的标签添加，帮助购物车主人建立数量大的目标概念标签。
由于图像识别引擎需要占用一定的计算机资源，因此可以预先为图像识别引擎申请服务器的计算资源。图像识别平台可以支持同时创建多个业务方分别对应的图像服务引擎。
5、响应于图像识别业务定制请求，建立目标业务标识与目标概念标签的对应关系。
具体地，当接收到图像识别业务定制请求，可以为目标业务方创建图像识别引擎，如创建对犬类进行识别的图像识别服务。如图12所示，由于图像识别基础模型已经预先训练得到，该图像识别基础模型可以输出标签树中各个实体标签对应的图像识别结果，因此可以基于购 物车中的目标概念标签以及已经训练得到的图像识别基础模型创建目标业务方对应的图像识别服务。
在一些实施例中，可以通过网页、母应用程序或者子应用程序的至少一种形式提供图像识别服务。其中，子应用程序俗称为小程序，是能够在母应用提供的环境中实现的应用程序。母应用程序是承载子应用程序的应用程序，为子应用程序的实现提供环境。母应用程序是原生应用程序。原生应用程序是可直接运行于操作系统的应用程序。例如，母应用程序可以是即时通讯应用程序，子应用程序可以是图像识别程序。当子应用程序发布后，可以在母应用程序对应的页面中展示子应用程序的图标，当用户点击该图标后，则可以在母应用提供的环境中运行该子应用程序，例如运行图像识别程序，从而使用户无需预先在终端中安装图像识别应用也可以进行图像识别。
6、当接收到携带目标业务标识的图像识别请求时，根据对应关系确定图像识别请求为针对目标概念标签的图像识别请求。
具体地，假设当接收到携带“犬类分类器”的图像识别请求时，根据“犬类分类器”可以得到对应的概念标签为“牧羊犬”以及“贵宾犬”，则确认该图像识别请求是针对“牧羊犬”以及“贵宾犬”这两个目标概念标签的图像识别请求。
7、将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，目标实体标签为目标概念标签对应的实体标签。
具体地，由于目标图像识别模型是多个概念标签对应的实体标签，所对应的训练图像训练得到的。因此该目标图像识别模型可以输出各个实体标签对应的概率。例如，可以输出“黄鹂鸟”、“燕子”、“德国牧羊犬”、“苏格兰牧羊犬”、“贵妇犬”、“卷毛狗”、“泰迪犬”分别对应的概率。其中，贵妇犬、卷毛狗、泰迪犬为贵宾犬的别称，即“贵妇犬”、“卷毛狗”、“泰迪犬”为概念标签“贵宾犬”对应的实体标签。因此目标概念标签“牧羊犬”对应的目标实体标签的图像识别结果包括“德国牧羊犬”以及“苏格兰牧羊犬”分别对应的概率。目标概念标签“贵宾犬”对应的目标实体标签的图像识别结果包括“贵妇犬”、“卷毛狗”、“泰迪犬”分别对应的概率。
8、根据目标图像识别结果集合得到目标概念标签对应的图像识别结果。
具体地，可以获取“德国牧羊犬”以及“苏格兰牧羊犬”分别对应的概率中最大的概率，作为目标概念标签“牧羊犬”对应的概率。可以获取“贵妇犬”、“卷毛狗”、“泰迪犬”分别对应的概率中最大的概率，作为目标概念标签“贵宾犬”对应的概率。因此，即使没有训练得到输出“牧羊犬”以及“贵宾犬”概念标签的识别概率的模型，也可以可得到“牧羊犬”以及“贵宾犬”概念标签的概率，相当于图像识别引擎模型可以推断得到在模型训练过程中不曾“见”过的概念标签。
由于不同的图像识别业务所需要输出的图像标签有着巨大的差别。比如游戏识别场景下的业务只关注游戏图像，并输出游戏相关的标签即可；而菜品识别场景的业务则关心输出的菜品标签是否齐全，是否正确。如果针对不同的业务需求单独训练该业务的图像识别模型， 并上线图像识别引擎是极其耗费精力、财力和资源的。而通过本申请实施例提供的图像识别方法，通过训练得到输出识别低层级的实体标签对应的识别结果的图像识别模型。对于概念标签，输入到实体标签对应的图像识别模型中进行识别，再根据实体标签对应的识别结果得到概念标签的识别结果，因此可以采用通用的图像识别模型进行图像的识别
通过本申请实施例提供的图像识别方法进行图像识别，为业务方定制符合业务需要的图像识别引擎，使得服务提供方无需为了某个特定图像识别业务进行专门的标签制定和图像数据挖掘并训练模型。业务方也可以自主以及快速地构建所需要的图像识别引擎服务，从而能够节省大量的人力和资源，使服务提供方能够专注于开发识别率更高的图像识别模型，减少响应各种业务需求，进行业务配置的时间，同时业务方能够专注于业务，在无需了解解图像识别的特定细节的情况下也能够得到满足业务需求的服务。
应该理解的是，虽然上述流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，上述流程图中的至少一部分步骤可以包括多个步骤或者多个阶段，这些步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤中的步骤或者阶段的至少一部分轮流或者交替地执行。
在一些实施例中，如图13所示，提供了一种图像识别装置，该装置可以采用软件模块或硬件模块，或者是二者的结合成为计算机设备的一部分，该装置具体包括：图像识别请求接收模块1302、目标图像获取模块1304、目标图像识别结果集合获取模块1306和结果获取模块1308，其中：
图像识别请求接收模块1302，用于接收针对目标概念标签的图像识别请求。
目标图像获取模块1304，用于获取图像识别请求对应的待识别的目标图像。
目标图像识别结果集合获取模块1306，用于将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，目标实体标签为目标概念标签对应的实体标签。
结果获取模块1308，用于根据目标图像识别结果集合得到目标概念标签对应的图像识别结果。
在一些实施例中，目标图像识别结果集合获取模块1306包括：候选图像识别结果集合获取单元，用于将目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到候选图像识别结果集合，候选图像识别结果集合包括第一实体标签对应的图像识别结果；目标实体标签确定单元，用于确定目标概念标签对应的第一实体标签，作为目标实体标签；目标图像识别结果集合获取单元，用于从候选图像识别结果集合中获取目标实体标签对应的图像识别结果，得到目标图像识别结果集合。
在一些实施例中，图像识别装置还包括：第一实体标签集合获取模块，用于获取第一实体标签集合，第一实体标签集合中包括多个概念标签分别对应的实体标签；训练图像集合得到模块，用于获取第一实体标签集合中各个第一实体标签分别对应的图像，作为训练图像， 得到训练图像集合；模型训练模块，用于根据训练图像集合进行模型训练，得到目标图像识别模型，将目标图像识别模型作为多个概念标签对应的通用图像识别模型。
在一些实施例中，目标实体标签对应的图像识别结果包括：目标图像中包括目标实体标签对应的实体的可能度，结果获取模块1308用于：从目标图像识别结果集合中获取满足可能度条件的可能度，作为目标可能度；根据目标可能度得到目标概念标签对应的图像识别结果，可能度条件包括可能度排序在第一排序之前、可能度排序为第一排序或者可能度大于第一预设阈值中的至少一个。
在一些实施例中，可能度条件为可能度最大，结果获取模块1308用于执行以下步骤的至少一个：当目标可能度大于第二预设阈值时，确定目标概念标签对应的图像识别结果包括：目标图像中包括目标概念标签对应的实体；或者将目标可能度作为目标概念标签对应的图像识别结果。
在一些实施例中，图像识别装置还包括：业务定制请求接收模块，用于接收目标业务终端发送的图像识别业务定制请求，图像识别业务定制请求携带目标业务标识以及对应的目标概念标签；对应关系建立模块，用于响应于图像识别业务定制请求，建立目标业务标识与目标概念标签的对应关系；图像识别请求接收模块用于：当接收到携带目标业务标识的图像识别请求时，根据对应关系确定图像识别请求为针对目标概念标签的图像识别请求。
在一些实施例中，图像识别装置还包括：标签协作请求接收模块，用于接收目标业务标识对应的标签协作请求；目标概念标签得到模块，用于根据在标签协作请求对应的协作终端上触发选择的概念标签，得到目标业务标识对应的目标概念标签。
在一些实施例中，目标概念标签得到模块用于：将标签协作请求对应的协作终端触发选择的概念标签，发送至目标业务终端的概念标签显示区域中，概念标签显示区域用于显示选择的概念标签；将目标业务终端接收到触发图像识别业务定制请求的触发操作时，概念标签显示区域当前显示的概念标签，作为目标业务标识对应的目标概念标签。
在一些实施例中，标签协作请求携带协作用户标识、目标业务标识以及协作标签范围，图像识别装置还包括：协作标签获取模块，用于获取协作标签范围对应的标签，作为协作标签；协作信息发送模块，用于将协作标签以及目标业务标识发送至协作用户标识对应的协作终端，以使得协作终端在标签协作选择界面上显示目标业务标识以及对应的协作标签。
关于图像识别装置的具体限定可以参见上文中对于图像识别方法的限定，在此不再赘述。上述图像识别装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中，提供了一种计算机设备，该计算机设备可以是服务器，其内部结构图可以如图14所示。该计算机设备包括通过系统总线连接的处理器、存储器和网络接口。其中，该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的 数据库用于存储图像。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种图像识别方法。
本领域技术人员可以理解，图14中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
在一些实施例中，还提供了一种计算机设备，包括存储器和处理器，存储器中存储有计算机可读指令，该处理器执行计算机可读指令时实现上述各方法实施例中的步骤。
在一些实施例中，提供了一种计算机可读存储介质，存储有计算机可读指令，该计算机可读指令被处理器执行时实现上述各方法实施例中的步骤。
在一些实施例中，提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述各方法实施例中的步骤。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机可读指令来指令相关的硬件来完成，所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中，该计算机可读指令在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和易失性存储器中的至少一种。非易失性存储器可包括只读存储器(Read-Only Memory，ROM)、磁带、软盘、闪存或光存储器等。易失性存储器可包括随机存取存储器(Random Access Memory，RAM)或外部高速缓冲存储器。作为说明而非局限，RAM可以是多种形式，比如静态随机存取存储器(Static Random Access Memory，SRAM)或动态随机存取存储器(Dynamic Random Access Memory，DRAM)等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (15)

  1. 一种图像识别方法，由计算机设备执行，所述方法包括：
    接收针对目标概念标签的图像识别请求；
    获取所述图像识别请求对应的待识别的目标图像；
    将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，所述目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，所述目标实体标签为所述目标概念标签对应的实体标签；以及
    根据所述目标图像识别结果集合得到所述目标概念标签对应的图像识别结果。
  2. 根据权利要求1所述的方法，其特征在于，所述将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合包括：
    将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到候选图像识别结果集合，所述候选图像识别结果集合包括第一实体标签对应的图像识别结果；
    确定所述目标概念标签对应的第一实体标签，作为目标实体标签；以及
    从所述候选图像识别结果集合中获取所述目标实体标签对应的图像识别结果，得到目标图像识别结果集合。
  3. 根据权利要求2所述的方法，其特征在于，所述方法还包括：
    获取第一实体标签集合，所述第一实体标签集合中包括多个概念标签分别对应的实体标签；
    获取所述第一实体标签集合中各个第一实体标签分别对应的图像，作为训练图像，得到训练图像集合；以及
    根据所述训练图像集合进行模型训练，得到所述目标图像识别模型，将所述目标图像识别模型作为所述多个概念标签对应的通用图像识别模型。
  4. 根据权利要求1所述的方法，其特征在于，所述目标实体标签对应的图像识别结果包括：所述目标图像中包括所述目标实体标签对应的实体的可能度，所述根据所述目标图像识别结果集合得到所述概念标签对应的图像识别结果包括：
    从所述目标图像识别结果集合中获取满足可能度条件的可能度，作为目标可能度；以及
    根据所述目标可能度得到所述目标概念标签对应的图像识别结果，所述可能度条件包括可能度排序在第一排序之前、可能度排序为所述第一排序或者可能度大于第一预设阈值中的至少一个。
  5. 根据权利要求4所述的方法，其特征在于，所述可能度条件为可能度最大，所述根据所述目标可能度得到所述目标概念标签对应的图像识别结果包括以下步骤的至少一个：
    当所述目标可能度大于第二预设阈值时，确定所述目标概念标签对应的图像识别结果包括：所述目标图像中包括所述目标概念标签对应的实体；或者
    将所述目标可能度作为所述目标概念标签对应的图像识别结果。
  6. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    接收目标业务终端发送的图像识别业务定制请求，所述图像识别业务定制请求携带目标 业务标识以及对应的所述目标概念标签；
    响应于所述图像识别业务定制请求，建立所述目标业务标识与所述目标概念标签的对应关系；
    所述接收针对目标概念标签的图像识别请求包括：
    当接收到携带所述目标业务标识的图像识别请求时，根据所述对应关系确定所述图像识别请求为针对所述目标概念标签的图像识别请求。
  7. 根据权利要求6所述的方法，其特征在于，所述方法还包括：
    接收所述目标业务标识对应的标签协作请求；以及
    根据在所述标签协作请求对应的协作终端上触发选择的概念标签，得到所述目标业务标识对应的目标概念标签。
  8. 根据权利要求7所述的方法，其特征在于，所述根据所述标签协作请求对应的协作终端触发选择的概念标签，得到所述目标业务标识对应的目标概念标签包括：
    将所述标签协作请求对应的协作终端触发选择的概念标签，发送至所述目标业务终端的概念标签显示区域中，所述概念标签显示区域用于显示已选择的概念标签；以及
    将所述目标业务终端接收到触发所述图像识别业务定制请求的触发操作时，所述概念标签显示区域当前显示的概念标签，作为所述目标业务标识对应的目标概念标签。
  9. 根据权利要求7所述的方法，其特征在于，所述标签协作请求携带协作用户标识、所述目标业务标识以及协作标签范围，所述方法还包括：
    获取所述协作标签范围对应的标签，作为协作标签；以及
    将所述协作标签以及所述目标业务标识发送至所述协作用户标识对应的协作终端，以使得所述协作终端在标签协作选择界面上显示所述目标业务标识以及对应的所述协作标签。
  10. 一种图像识别装置，所述装置包括：
    图像识别请求接收模块，用于接收针对目标概念标签的图像识别请求；
    目标图像获取模块，用于获取所述图像识别请求对应的待识别的目标图像；
    目标图像识别结果集合获取模块，用于将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到目标图像识别结果集合，所述目标图像识别结果集合中的图像识别结果为目标实体标签对应的图像识别结果，所述目标实体标签为所述目标概念标签对应的实体标签；以及
    结果获取模块，用于根据所述目标图像识别结果集合得到所述目标概念标签对应的图像识别结果。
  11. 根据权利要求10所述的装置，其特征在于，所述目标图像识别结果集合获取模块包括：
    候选图像识别结果集合获取单元，用于将所述目标图像输入到实体标签对应的目标图像识别模型中进行图像识别，得到候选图像识别结果集合，所述候选图像识别结果集合包括第一实体标签对应的图像识别结果；
    目标实体标签确定单元，用于确定所述目标概念标签对应的第一实体标签，作为目标实 体标签；以及
    目标图像识别结果集合获取单元，用于从所述候选图像识别结果集合中获取所述目标实体标签对应的图像识别结果，得到目标图像识别结果集合。
  12. 根据权利要求11所述的装置，其特征在于，所述装置还包括：
    第一实体标签集合获取模块，用于获取第一实体标签集合，所述第一实体标签集合中包括多个概念标签分别对应的实体标签；
    训练图像集合得到模块，用于获取所述第一实体标签集合中各个第一实体标签分别对应的图像，作为训练图像，得到训练图像集合；以及
    模型训练模块，用于根据所述训练图像集合进行模型训练，得到所述目标图像识别模型，将所述目标图像识别模型作为所述多个概念标签对应的通用图像识别模型。
  13. 根据权利要求10所述的装置，其特征在于，所述目标实体标签对应的图像识别结果包括：所述目标图像中包括所述目标实体标签对应的实体的可能度，所述结果获取模块用于：
    从所述目标图像识别结果集合中获取满足可能度条件的可能度，作为目标可能度；以及
    根据所述目标可能度得到所述目标概念标签对应的图像识别结果，所述可能度条件包括可能度排序在第一排序之前、可能度排序为第一排序或者可能度大于第一预设阈值中的至少一个。
  14. 一种计算机设备，其特征在于，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述计算机可读指令被所述处理器执行时实现权利要求1至9中任一项所述方法的步骤。
  15. 一个或多个存储有计算机可读指令的非易失性存储介质，所述计算机可读指令被一个或多个处理器执行时实现权利要求1至9中任一项所述的方法的步骤。
PCT/CN2020/127455 2020-03-18 2020-11-09 图像识别方法、装置、计算机设备和存储介质 WO2021184776A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010189599.3 2020-03-18
CN202010189599.3A CN111368789B (zh) 2020-03-18 2020-03-18 图像识别方法、装置、计算机设备和存储介质

Publications (1)

Publication Number Publication Date
WO2021184776A1 true WO2021184776A1 (zh) 2021-09-23

Family

ID=71212530

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/127455 WO2021184776A1 (zh) 2020-03-18 2020-11-09 图像识别方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN111368789B (zh)
WO (1) WO2021184776A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115061490A (zh) * 2022-05-30 2022-09-16 广州中科云图智能科技有限公司 基于无人机的水库巡检方法、装置、设备以及存储介质

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111368789B (zh) * 2020-03-18 2023-05-26 腾讯科技（深圳）有限公司 图像识别方法、装置、计算机设备和存储介质
CN112380261B (zh) * 2020-10-10 2024-06-18 杭州翔毅科技有限公司 基于5g技术的远程辅导方法、装置、系统及存储介质
US11769256B2 (en) * 2021-02-16 2023-09-26 Avanade Holdings Llc Image creation for computer vision model training
CN113298806A (zh) * 2021-06-18 2021-08-24 上海市机械施工集团有限公司 焊接缺陷智能检测方法、装置、焊接设备和存储介质
CN115203457B (zh) * 2022-07-15 2023-11-14 小米汽车科技有限公司 图像检索方法、装置、车辆、存储介质及芯片

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120027288A1 (en) * 2010-07-29 2012-02-02 Chang Yuan Methods, Systems and Apparatus for Defect Detection
CN104200238A (zh) * 2014-09-22 2014-12-10 北京酷云互动科技有限公司 台标识别方法和台标识别装置
CN110084317A (zh) * 2019-05-06 2019-08-02 北京字节跳动网络技术有限公司 用于识别图像的方法和装置
CN110414532A (zh) * 2019-06-21 2019-11-05 广州利科科技有限公司 一种基于深度学习的商品图像识别方法
CN110852360A (zh) * 2019-10-30 2020-02-28 腾讯科技（深圳）有限公司 图像情感识别方法、装置、设备及存储介质
CN111368789A (zh) * 2020-03-18 2020-07-03 腾讯科技（深圳）有限公司 图像识别方法、装置、计算机设备和存储介质

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104751163B (zh) * 2013-12-27 2018-06-19 同方威视技术股份有限公司 对货物进行自动分类识别的透视检查系统和方法
US10019654B1 (en) * 2017-06-28 2018-07-10 Accenture Global Solutions Limited Image object recognition
CN108171257B (zh) * 2017-12-01 2019-11-26 百度在线网络技术（北京）有限公司 细粒度图像识别模型训练及识别方法、装置及存储介质
CN109987102A (zh) * 2018-01-03 2019-07-09 田文洪 一种基于卷积神经网络的高精度自动识别驾驶员不安全行为的方法与装置
CN109584050A (zh) * 2018-12-14 2019-04-05 深圳壹账通智能科技有限公司 基于微表情识别的用户风险程度分析方法及装置
CN110163301A (zh) * 2019-05-31 2019-08-23 北京金山云网络技术有限公司 一种图像的分类方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120027288A1 (en) * 2010-07-29 2012-02-02 Chang Yuan Methods, Systems and Apparatus for Defect Detection
CN104200238A (zh) * 2014-09-22 2014-12-10 北京酷云互动科技有限公司 台标识别方法和台标识别装置
CN110084317A (zh) * 2019-05-06 2019-08-02 北京字节跳动网络技术有限公司 用于识别图像的方法和装置
CN110414532A (zh) * 2019-06-21 2019-11-05 广州利科科技有限公司 一种基于深度学习的商品图像识别方法
CN110852360A (zh) * 2019-10-30 2020-02-28 腾讯科技（深圳）有限公司 图像情感识别方法、装置、设备及存储介质
CN111368789A (zh) * 2020-03-18 2020-07-03 腾讯科技（深圳）有限公司 图像识别方法、装置、计算机设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115061490A (zh) * 2022-05-30 2022-09-16 广州中科云图智能科技有限公司 基于无人机的水库巡检方法、装置、设备以及存储介质
CN115061490B (zh) * 2022-05-30 2024-04-05 广州中科云图智能科技有限公司 基于无人机的水库巡检方法、装置、设备以及存储介质

Also Published As

Publication number Publication date
CN111368789A (zh) 2020-07-03
CN111368789B (zh) 2023-05-26

Similar Documents

Publication Publication Date Title
WO2021184776A1 (zh) 图像识别方法、装置、计算机设备和存储介质
CN109919316B (zh) 获取网络表示学习向量的方法、装置和设备及存储介质
CN109447140B (zh) 一种基于神经网络深度学习的图像识别并推荐认知的方法
CN111431742B (zh) 网络信息检测方法、装置、存储介质和计算机设备
US11645548B1 (en) Automated cloud data and technology solution delivery using machine learning and artificial intelligence modeling
CN111125422A (zh) 一种图像分类方法、装置、电子设备及存储介质
CN111353091A (zh) 信息处理方法、装置、电子设备及可读存储介质
KR102607208B1 (ko) 뉴럴 네트워크 학습 방법 및 디바이스
CN112256537B (zh) 模型运行状态的展示方法、装置、计算机设备和存储介质
CN111885399A (zh) 内容分发方法、装置、电子设备以及存储介质
US11907860B2 (en) Targeted data acquisition for model training
CN108549909B (zh) 基于众包的对象分类方法及对象分类系统
JP6787831B2 (ja) 検索結果による学習が可能な対象検出装置、検出モデル生成装置、プログラム及び方法
US20240211496A1 (en) Systems and Methods for Determining Entity Attribute Representations
US20190228270A1 (en) Image analysis enhanced related item decision
US20230106106A1 (en) Text backup method, apparatus, and device, and computer-readable storage medium
CN113244627B (zh) 识别外挂的方法、装置、电子设备以及存储介质
US20230119860A1 (en) Matching system, matching method, and matching program
CN114091572A (zh) 模型训练的方法、装置、数据处理系统及服务器
CN113704534A (zh) 图像处理方法、装置及计算机设备
US20230117893A1 (en) Machine learning techniques for environmental discovery, environmental validation, and automated knowledge repository generation
Sinnott et al. Run or pat: using deep learning to classify the species type and emotion of pets
Qu et al. The application of interactive methods under swarm computing and artificial intelligence in image retrieval and personalized analysis
CN117033664B (zh) 业务的序列图生成方法、装置、计算机设备和存储介质
US12041099B2 (en) Data modeling for virtual collaboration environment

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: 20925759

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22/02/2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20925759

Country of ref document: EP

Kind code of ref document: A1