Electronic device and control method thereof

Abstract

An electronic device is disclosed. The electronic device according to the disclosure includes a processor, wherein the processor obtains a photographed image, inputs the photographed image into a first neural network model to obtain a feature value of an object included in the photographed image, a predicted category, and a probability value of the predicted category, identifies a training image from the photographed image based on the probability value, obtains training data by clustering the feature value of the object included in the training image, and receives information about a second neural network model obtained based on the training data from an external device, thereby making it possible to update the first neural network model.

Description

Technical Field

[0001] This disclosure relates to an electronic device and a control method thereof, and more specifically, to an electronic device and a control method thereof for updating a neural network model. Background Technology

[0002] Recently, artificial intelligence (AI) systems have been applied in various fields. Unlike traditional rule-based intelligent systems, AI systems are systems where machines learn, determine, and become more intelligent on their own. As AI systems are used more extensively, they demonstrate improved recognition rates and become more accurately able to understand user preferences. Therefore, traditional rule-based intelligent systems are gradually being replaced by deep learning-based AI systems.

[0003] Artificial intelligence technology consists of machine learning (e.g., deep learning) and the underlying technologies that utilize machine learning.

[0004] Machine learning refers to algorithmic techniques that classify / learn the characteristics of input data. Fundamental technologies, on the other hand, refer to techniques that use machine learning algorithms (such as deep learning) to perform human brain functions (such as cognition and determination), and comprise technical fields such as language understanding, visual understanding, inference / prediction, knowledge representation, and operational control. Specifically, language understanding refers to techniques for recognizing and applying / processing human language / characters, and includes natural language processing, machine translation, communication systems, query and response, speech recognition / synthesis, etc.

[0005] Recently, various electronic devices (e.g., robotic cleaners) have been released, including those employing neural network models for object recognition. However, due to performance limitations such as the CPU or memory capacity of electronic devices, the neural network models stored within them cannot classify all objects. Generally, these models are trained to classify only objects pre-specified by the manufacturer. Therefore, there is an increasing demand from users to update neural network models to better suit the operating environment of electronic devices.

[0006] Therefore, the demand for technologies to update neural network models is increasing. Summary of the Invention

[0007] Technical issues

[0008] The technical task to be addressed by this disclosure is to provide an electronic device for obtaining learning data for updating a neural network model.

[0009] Furthermore, the technical tasks of this disclosure are not limited to those described above, and other technical tasks not mentioned will be clearly understood by those skilled in the art to which this disclosure pertains based on the following description.

[0010] Technical solution

[0011] According to one aspect of this disclosure, a method for controlling an electronic device is provided, the method comprising: acquiring a plurality of images, wherein the plurality of images include one or more objects; inputting the plurality of images into a first neural network model for recognizing the objects, and based on the input of the plurality of images into the first neural network model for recognizing the objects, obtaining feature values ​​of each of the one or more objects, a predicted category of each of the one or more objects based on the corresponding obtained feature values, and a probability value of the predicted category of each of the one or more objects; recognizing one or more learning images from the plurality of images based on the obtained probability values; recognizing one or more clusters of feature values ​​by mapping the feature values ​​of the one or more objects included in the one or more recognized learning images to a vector space; obtaining learning data from the one or more recognized learning images based on the obtained feature values; sending the obtained learning data to an external device; receiving information about a second neural network model from the external device; and updating the first neural network model based on the received information about the second neural network model.

[0012] The step of identifying one or more learning images may further include: identifying one or more learning images based on objects whose probability values ​​are less than a predetermined value among one or more objects.

[0013] The steps of obtaining learning data may further include: identifying at least one cluster among a plurality of clusters of feature values ​​that include feature values ​​of one or more objects included in one or more identified learning images, and having a cohesion greater than a predetermined value; and obtaining one or more images corresponding to the feature values ​​included in the at least one cluster of identified feature values ​​as learning data.

[0014] The step of obtaining one or more images corresponding to feature values ​​included in at least one cluster of the identified feature values ​​as learning data may further include: identifying the learning data and obtaining images corresponding to the average closest feature values ​​included in at least one cluster of the identified feature values.

[0015] The method for controlling the electronic device may further include: storing the acquired learning data, wherein storing the acquired learning data may include: identifying clusters of feature values ​​that include a predetermined number of feature values ​​in at least one cluster of identified feature values, and storing images corresponding to the feature values ​​included in the identified clusters.

[0016] The learning data may be at least one of one or more learning images, and the learning data includes positional information of pixels corresponding to objects included in the at least one image.

[0017] The second neural network model can be a model obtained based on a third neural network that has higher accuracy than the first neural network, and the third neural network is trained based on the obtained learning data.

[0018] According to one aspect of this disclosure, an electronic device is provided, comprising: a communication interface including at least one circuit; a memory including at least one instruction; and a processor. The processor is configured to: acquire a plurality of images, wherein the plurality of images include one or more objects; input the plurality of images into a first neural network model for recognizing the objects, and based on the input of the plurality of images into the first neural network model for recognizing the objects, obtain feature values ​​for each of the one or more objects, a predicted category for each of the one or more objects based on the corresponding obtained feature values, and a probability value for the predicted category of each of the one or more objects; recognize one or more learning images from the plurality of images based on the obtained probability values; recognize one or more clusters of feature values ​​by mapping the feature values ​​of one or more objects included in the one or more recognized learning images to a vector space; obtain learning data from the one or more recognized learning images based on the obtained feature values; transmit the obtained learning data to an external device via the communication interface; receive information about a second neural network model from the external device; and update the first neural network model based on the received information about the second neural network model.

[0019] The processor can also be configured to identify one or more learning images based on objects whose probability values ​​are less than a predetermined value, including one or more objects.

[0020] The processor can also be configured to: identify at least one cluster with a cohesion greater than a predetermined value among a plurality of clusters of feature values ​​including feature values ​​of one or more objects included in one or more identified learning images; and obtain one or more images corresponding to feature values ​​included in at least one cluster of identified feature values ​​as learning data.

[0021] The processor can also be configured to: identify and obtain an image as learning data corresponding to the average closest feature value included in at least one cluster of the identified feature values.

[0022] The processor can also be configured to: identify clusters of feature values ​​that include a predetermined number of feature values ​​in at least one cluster of identified feature values, and store images corresponding to the feature values ​​included in the identified clusters.

[0023] The learning data may be at least one of one or more learning images, and the learning data may include positional information of pixels corresponding to objects included in the at least one image.

[0024] The second neural network model can be a model obtained based on a third neural network that has higher accuracy than the first neural network, and the third neural network is trained based on the obtained learning data.

[0025] According to one aspect of this disclosure, a neural network model update system including an electronic device and an external device is provided. The electronic device can be configured to: acquire multiple images, obtain learning data from the multiple images using a first neural network model, send the acquired learning data to an external device, receive information about a second neural network model from the external device, and update the first neural network model based on the information about the second neural network model obtained from the external device. The external device can be configured to: input the learning data into a fourth neural network model and obtain labeled data, train a third neural network model based on the labeled data, obtain information about the second neural network model by compressing the third neural network model, and send the information about the second neural network model to the electronic device.

[0026] Furthermore, the means for solving the tasks of this disclosure are not limited to those described above, and other devices, systems, and methods not mentioned will be clearly understood by those skilled in the art to which this disclosure pertains based on this specification and the accompanying drawings.

[0027] Invention Effects

[0028] According to the various embodiments of this disclosure described above, an electronic device can update a neural network model stored in the electronic device. Therefore, the object recognition rate of the neural network model can be improved.

[0029] In addition to the foregoing description, effects that can be obtained or predicted from the embodiments of this disclosure will be described directly or implicitly in the detailed description of the embodiments of this disclosure. For example, various effects predicted according to the embodiments of this disclosure will be described in the specific implementation described below. Attached Figure Description

[0030] Figure 1 This is a diagram illustrating a neural network model update system according to an embodiment of the present disclosure;

[0031] Figure 2 This is a block diagram illustrating the configuration of an electronic device according to an embodiment of the present disclosure;

[0032] Figure 3 This is a diagram illustrating the configuration of an external device according to an embodiment of the present disclosure;

[0033] Figure 4A This is a diagram illustrating a method for recognizing and learning images according to an embodiment of the present disclosure;

[0034] Figure 4B This is a diagram illustrating a clustering method according to an embodiment of the present disclosure;

[0035] Figure 4C This is a diagram illustrating a method for obtaining learning data of an electronic device according to an embodiment of the present disclosure;

[0036] Figure 5A This is a diagram illustrating a method for recognizing and learning images according to another embodiment of the present disclosure;

[0037] Figure 5B This is a diagram illustrating a clustering method according to another embodiment of the present disclosure;

[0038] Figure 6A This is a diagram illustrating a method for obtaining marker data according to an embodiment of the present disclosure;

[0039] Figure 6B This is a diagram illustrating a method for obtaining a second neural network model according to an embodiment of the present disclosure;

[0040] Figure 6C This is a diagram illustrating a method for compressing a neural network model according to an embodiment of the present disclosure;

[0041] Figure 7A This is a flowchart illustrating a control method for an electronic device according to an embodiment of the present disclosure;

[0042] Figure 7B This is a flowchart illustrating a control method for an electronic device according to another embodiment of the present disclosure; and

[0043] Figure 8 This is a sequence diagram illustrating a neural network model update system according to an embodiment of the present disclosure. Detailed Implementation

[0044] First, the terminology used in this specification is briefly described, and then this disclosure is described in detail.

[0045] In consideration of the functions described in this disclosure, widely used and common terms have been selected as much as possible for use in the embodiments of this disclosure. However, these terms may vary depending on the intent of those skilled in the art, previous court judgments, or the emergence of new technologies. Furthermore, in certain cases, terms may be specified by the applicant, and in such cases, the meaning of the terms will be described in detail in the relevant descriptions of this disclosure. Therefore, the terms used in this disclosure should be defined based on their meaning and the overall content of this disclosure, and not merely on their names.

[0046] Furthermore, various modifications can be made to the embodiments of this disclosure, and various types of embodiments are possible. Therefore, specific embodiments will be illustrated in the accompanying drawings, and these embodiments will be described in detail in the specific implementation. However, it should be noted that the various embodiments are not intended to limit the scope of this disclosure to the specific embodiments, but rather they should be construed as including all modifications, equivalents, or substitutions of the embodiments included within the scope of the ideas and techniques disclosed herein.

[0047] Furthermore, terms such as "first" and "second" can be used to describe various components, but components should not be limited by these terms. These terms are only used to distinguish one component from another.

[0048] Furthermore, singular expressions include plural expressions, provided they do not have a clearly different meaning in the context. Additionally, in this disclosure, terms such as "comprising" and "including" should be interpreted as specifying the features, numbers, steps, operations, elements, components, or combinations thereof described in the specification, but do not preclude the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or the possibility of adding one or more other features, numbers, steps, operations, elements, components, or combinations thereof. Embodiments of this disclosure will now be described in detail with reference to the accompanying drawings to enable those skilled in the art to readily implement this disclosure. However, it should be noted that this disclosure can be implemented in various different forms and is not limited to the embodiments described herein. Furthermore, in the drawings, portions irrelevant to the explanation have been omitted for clarity, and similar components are designated by similar reference numerals throughout the specification.

[0049] Figure 1This diagram illustrates a neural network model update system according to an embodiment of the present disclosure. The neural network model update system may include an electronic device 100 and an external device 200. The electronic device 100 is a device for storing a neural network model, and may be, for example, a robotic cleaner, a smart robot, or a smart camera. The external device 200 is a device with higher performance than the electronic device 100, and may have faster processing speed and larger memory capacity compared to the electronic device 100. For example, the external device 200 may be a personal computer (PC) or a mobile terminal device. The neural network model update system can update the neural network model stored in the electronic device 100 by using the external device 200, which has relatively higher performance than the electronic device 100. The method for updating the neural network model will be described in more detail below.

[0050] The electronic device 100 can acquire multiple captured images 10. Here, the electronic device 100 can acquire multiple captured images 10 by using a camera 110 mounted on the electronic device 100 to capture images of the area around the electronic device 100.

[0051] Electronic device 100 can obtain an output value 20 as an inference result for object ob by inputting multiple captured images 10 into a first neural network model 41 stored in memory 140. Here, the first neural network model 41 is a neural network model trained to recognize or identify objects included in an image. Furthermore, the output value 20 is the inference result of the first neural network model 41 for object ob, and the output value 20 may include the classification of object ob and a probability value for that classification value. The classification value may refer to the predicted category of object ob. The probability value is a value indicating the probability that object ob is the object corresponding to that classification value; that is, a value indicating the probability that the predicted category of object ob is the actual category of object ob, and this probability value may also be referred to as a score. For example, the output value 20 of the first neural network model 41 for object ob may be "{1 (value corresponding to cat), 0.6 (probability that the object is a cat)}". Furthermore, the output value 20 may include information about the region where object ob is located within the captured image 10. For example, the information about the region where object ob is located may be the coordinate information of the object's bounding box.

[0052] Meanwhile, due to performance limitations of the electronic device 100, such as CPU or memory capacity, the first neural network model 41 cannot classify all objects, and generally, the neural network model is trained to classify only objects pre-specified by the manufacturer. Therefore, the first neural network model 41 needs to be updated to be more suitable for the environment in which the electronic device 100 is located. To this end, the electronic device 100 can obtain learning data 30 for updating the first neural network model 41.

[0053] Specifically, the electronic device 100 can identify an image of a plurality of captured images 10 that includes an object with a probability value smaller than a predetermined value (e.g., 0.7) as a learning image 15 (hereinafter referred to as a "learning image"). For example, if the probability value of object ob obtained by the first neural network model 41 is less than the predetermined value, the electronic device 100 can identify an image including object ob as the learning image 15.

[0054] Then, the electronic device 100 can obtain the feature values ​​V of the objects included in the learning image 15. n Here, the eigenvalue V n These are values ​​that indicate the characteristics of an object, and these characteristic values ​​can also be called feature vectors or feature information. Electronic device 100 can acquire feature values ​​V of objects included in multiple captured images 10. n As an intermediate output value, this intermediate output value is the result of inputting multiple captured images 10 into the first neural network model 41. Feature value V n It is the value of the characteristic of the object ob, and this characteristic value V n It can also be called a feature vector or feature information. Furthermore, the eigenvalue V n This could correspond to an output value of 20. Specifically, it could be based on the eigenvalue v. n To obtain the output value 20.

[0055] Electronic device 100 can obtain the characteristic value V n Clustering is performed by mapping to a random vector space. Clustering refers to grouping feature values ​​with similar characteristics together by mapping feature values ​​to a random vector space. Electronic device 100 can identify clustering feature values ​​V in the vector space. n Electronic device 100 can be based on the identified feature value V n Learning data 30 is obtained from multiple learning images 15. Specifically, the electronic device 100 can identify cluster C among multiple clusters existing in the vector space that has a cohesion greater than a predetermined value. n Here, cohesion refers to an index indicating the degree of cohesion among related feature information in a vector space. Electronic device 100 can identify clusters C... n The feature value that is closest to the average of the multiple feature values ​​included in the cluster is identified. The electronic device 100 can identify the learning image 15 corresponding to the feature value included in the identified cluster as learning data 30.

[0056] Then, the electronic device 100 can send the acquired learning data 30 to the external device 200. The external device 200 can obtain a second neural network model 42 based on the learning data 30. Specifically, the external device 200 can train a third neural network model 43 with higher performance than the first neural network model 41 based on the learning data 30. For example, the third neural network model 43 can identify more types of objects than the first neural network model 41. Then, the external device 200 can reduce the weights of the third neural network model 43 and obtain the second neural network model 42. Meanwhile, [the following will be implemented / conducted / etc.]. Figure 3 The process by which the external device 200 obtains the second neural network model 42 is described in more detail.

[0057] External device 200 can send information about the obtained second neural network model 42 to electronic device 100. Here, the information about the second neural network model 42 can be information about the second neural network model 42 itself, or parameters included in the second neural network model 42. Parameters can be weighted values. Weighted values ​​can be vector values ​​comprising several elements.

[0058] The electronic device 100 can update the first neural network model 41 based on information about the second neural network model 42. For example, the electronic device 100 can store the second neural network model 42 in the memory 140 and delete the first neural network model 41. Alternatively, the electronic device 100 can change the parameters of the first neural network model 41 to the parameters of the second neural network model 42.

[0059] The preceding text described the operation of the neural network model update system. The following text will describe the configuration of the electronic and external devices that constitute the neural network model update system.

[0060] Figure 2 This is a block diagram illustrating the configuration of an electronic device according to an embodiment of the present disclosure. Reference Figure 2 The electronic device 100 may include a camera 110, a communication interface 120, a drive component 130, a memory 140, and a processor 150.

[0061] Camera 110 can acquire multiple captured images. For example, camera 110 can be mounted on electronic device 100 and capture images of the area around electronic device 100. Furthermore, camera 110 can be implemented as various types of cameras. For example, camera 110 can be implemented as a 2D-based RGB camera or IR camera, or as a 3D-based Time-of-Flight (ToF) camera or stereo camera. Additionally, camera 110 can be implemented as a combination of two or more of the aforementioned camera types.

[0062] The communication interface 120 includes at least one circuit and can communicate with various types of external devices according to various communication methods. The communication interface 120 may include a Wi-Fi chip and a Bluetooth chip. The electronic device 100 can communicate with an external server or external device 200 through the communication interface 120.

[0063] The drive component 130 may be a component for moving the electronic device 100. In particular, the drive component 130 may include an actuator for driving the electronic device 100. In addition, besides the drive component 130, it may also include actuators for driving the movement of other physical components of the electronic device 100 (e.g., an arm, etc.).

[0064] Memory 140 may store the operating system (OS) for controlling the overall operation of the components of electronic device 100, as well as instructions or data related to the components of electronic device 100. For this purpose, memory 140 may be implemented as non-volatile memory (e.g., hard disk, solid-state drive (SSD), flash memory), volatile memory, etc. Memory 140 may store neural network models for identifying or recognizing objects. In particular, the neural network model may be executed by a conventional general-purpose processor (e.g., CPU) or a dedicated AI processor (e.g., GPU, NPU, etc.). For example, memory 140 may include a first neural network model 41 for recognizing objects. Furthermore, the first neural network model 41 may be updated to a second neural network model 42, as described above. Figure 1 As described in [the text].

[0065] The processor 150 can control the overall operation of the electronic device 100. The processor 150 may include an object inference module 151, a learning image recognition module 152, a learning data acquisition module 153, a learning data storage module 154, and a neural network model update module 155. The operation of each module will be described in detail below.

[0066] The object inference module 151 can obtain the feature values ​​V of objects included in multiple captured images 10 obtained by the camera 110. n Furthermore, the object inference module 151 can obtain classification values ​​(i.e., predicted categories) and probability values ​​corresponding to each feature value.

[0067] The learning image recognition module 152 can identify the learning image 15 from multiple captured images 10 based on the probability value of the object ob obtained by the object inference module 151. Specifically, the learning image recognition module 152 can identify captured images 10 that include objects ob with probability values ​​less than a predetermined value as the learning image 15.

[0068] The learning data acquisition module 153 can obtain learning data 30 based on the learning image 15. Specifically, the learning data acquisition module 153 can obtain learning data 30 by taking the feature values ​​V of the object ob included in the learning image 15. n The feature values ​​are clustered by mapping to a random vector space. Then, the learning data acquisition module 153 can identify at least one cluster C among multiple clusters existing in the vector space that has a cohesion greater than a predetermined value. n Furthermore, the learning data acquisition module 153 can identify clusters with cohesion greater than a predetermined value that include a predetermined number of feature values ​​V. n At least one cluster.

[0069] The learning data acquisition module 153 can compare the identified clusters C n The learning image 15 corresponding to the feature values ​​included in the cluster is identified as learning data 30. Here, the learning data acquisition module 153 can identify the feature value V that is closest to the average of the multiple feature values ​​included in the identified cluster. n Alternatively, the learning data acquisition module 153 can identify and recognize the clusters C identified. n The eigenvalue V closest to the center n Simultaneously, the learning data acquisition module 153 can identify multiple feature values ​​within a cluster. Here, the learning data acquisition module 153 can identify features at a distance V from the feature value V. n Multiple feature values ​​V within the average threshold range n Alternatively, the learning data acquisition module 153 can identify multiple feature values ​​V within a range of the cluster center threshold. n .

[0070] Then, the learning data acquisition module 153 can include the identified feature values ​​V n The corresponding learning image 15 of the object ob is identified as learning data 30. Here, learning data 30 may include raw data about the learning image 15, as well as the position information of the object ob included in the learning image 15 that corresponds to the identified feature information. Here, the position information of the object ob may be the coordinate information of the bounding box of the object ob.

[0071] The learning data storage module 154 can store the acquired learning data 30 in the memory 140. Here, considering the capacity of the memory 140, the learning data storage module 154 can select only some of the learning data 30 and store it in the memory 140. Specifically, the learning data storage module 154 can identify the learning data 30 to be stored based on the priority of the clusters corresponding to the acquired learning data 30. Here, it can be based on the cluster C... n The eigenvalues ​​V includedn To set the number of clusters C n Priority. Furthermore, after obtaining the same cluster C... n In the case of multiple corresponding learning data 30, the learning data storage module 154 can be based on the feature values ​​V corresponding to the multiple corresponding learning data 30. n To select some from the learning data 30. For example, when the feature value V n Closer to the same cluster C n When the average of multiple feature values ​​included in the data is obtained, the learning data storage module 154 can store the feature value V. n The priority is identified as higher priority, and will be compared with the feature value V based on the identified priority. n The corresponding learning data 30 is stored in memory 140. Alternatively, when the feature value V n Closer to the same cluster C n When the center is reached, the learning data storage module 154 can store the feature value V. n The priority is identified as higher priority, and will be compared with the feature value V based on the identified priority. n The corresponding learning data 30 is stored in memory 140.

[0072] The neural network model update module 155 can update the first neural network model 41 based on information received from the external device 200 regarding the second neural network model 42. For example, the neural network model update module 155 can store the second neural network model 42 in the memory 140 and delete the first neural network model 41. Alternatively, the neural network model update module 155 can change the parameters (or weights) of the first neural network model 41 to the parameters of the second neural network model 42.

[0073] Specifically, the AI-related functions according to this disclosure are operated via processor 150 and memory 140. Processor 150 may consist of one or more processors. Here, the one or more processors 150 may be general-purpose processors (such as CPUs, APs, digital signal processors (DSPs), etc.), graphics-specific processors (such as GPUs and vision processing units (VPUs)), or AI-specific processors, such as NPUs. The one or more processors 150 may perform control to process input data according to predefined operating rules or AI models stored in memory 140. Alternatively, if the one or more processors 150 are AI-specific processors, the AI-specific processors may be designed with a hardware architecture specifically designed to process a particular AI model.

[0074] The characteristic of predefined operating rules or artificial intelligence models is that they are formed through learning. Here, "formed through learning" means that a predefined operating rule or artificial intelligence model is formed by training a basic artificial intelligence model using a learning algorithm and multiple learning data, thereby setting it to perform the desired characteristics (or purpose). This learning can be performed in the device itself, in which artificial intelligence is performed according to this disclosure, or via a separate server and / or system. As examples of learning algorithms, there are supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but the learning algorithms in this disclosure are not limited to the foregoing examples.

[0075] Artificial intelligence models can be formed through learning. Here, "formed through learning" refers to training a basic AI model using a learning algorithm and multiple learning datasets, thereby forming a predefined set of operational rules or an AI model configured to perform desired characteristics (or objectives). An AI model can consist of multiple neural network layers. Each of these layers has multiple weight values, and neural network operations are performed through operations between the results of the previous layer and these weight values. The multiple weight values ​​comprised of the neural network layers can be optimized based on the learning results of the AI ​​model. For example, multiple weight values ​​can be updated to reduce or minimize the loss or cost values ​​obtained at the AI ​​model during the learning process.

[0076] Visual understanding is a technology that identifies and processes objects in a way similar to human vision, and includes object recognition, object tracking, image retrieval, human recognition, scene recognition, spatial recognition (3D reconstruction / localization), image enhancement, etc.

[0077] Artificial neural networks may include deep neural networks (DNNs), and include, for example, convolutional neural networks (CNNs), deep neural networks (DNNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), restricted Boltzmann machines (RBMs), deep belief networks (DBNs), bidirectional recurrent deep neural networks (BRDNNs), or deep Q-networks, but this disclosure is not limited to the foregoing examples.

[0078] Figure 3 This is a diagram illustrating the configuration of an external device 200 according to an embodiment of the present disclosure. Reference Figure 3 The external device 200 may include a communication interface 210, a memory 220, and a processor 230. Meanwhile, the communication interface 210 may correspond to... Figure 2 The communication interface 120 is used, so a detailed explanation of it will be omitted.

[0079] The memory 220 can store the third neural network model 43 and the fourth neural network model 44. Here, the third neural network model 43 can be... Figure 2The first neural network model 41 is the original model. That is, the first neural network model 41 can be a weight-compressed version of the third neural network model 43. Furthermore, the fourth neural network model 44 can be a model with higher performance than the third neural network model 43, and can be a model trained to recognize a greater number of objects than the third neural network model 43. Additionally, the memory 220 can store the operating system (OS) for controlling the overall operation of the components of the external device 200, as well as instructions or data related to the components of the external device 200. For this purpose, the memory 220 can be implemented as non-volatile memory (e.g., hard disk, solid-state drive (SSD), flash memory), volatile memory, etc.

[0080] The processor 230 can control the overall operation of the external device 200. The processor 230 may include a learning data inference module 231, a neural network model training module 232, and a neural network model compression module 233. The operation of each module will be described in detail below.

[0081] The learning data inference module 231 can input the learning data 30 obtained from the electronic device 100 into the fourth neural network model 44 to obtain the classification values ​​and probability values ​​of the objects included in the learning data 30. Since the fourth neural network model 44 has higher performance than the first neural network model 41, the learning data inference module 231 can use the fourth neural network model 44 to obtain the classification values ​​and probability values ​​of objects ob that the first neural network model 41 cannot recognize. Then, the learning data inference module 231 can recognize objects ob with probability values ​​greater than a predetermined probability value (e.g., 0.7) and perform a labeling operation on the learning data 30 including the identified objects ob. That is, the learning data inference module 231 can match the learning data 30 with the classification values ​​of the objects ob included in the learning data 30 and store them in the memory 220. In the following text, the learning data 30 labeled by the learning data inference module 231 will be referred to as labeled data 31.

[0082] The neural network model training module 232 can train the third neural network model 43 based on the labeled data 31. Here, the neural network model training module 232 can train the third neural network model 43 based on the labeled data 31 and the learning data 30 of the first neural network model 41. Therefore, bias or overfitting of the third neural network model 43 can be prevented. Furthermore, since the neural network model training module 232 trains the third neural network model 43 (which is the original model of the first neural network model 41 (i.e., the model before compression)) instead of the first neural network model 41, the learning time can be shortened.

[0083] The neural network model compression module 233 can obtain the second neural network model 42 by compressing the weights of the trained third neural network model 43. Specifically, the neural network model compression module 233 can obtain the second neural network model 42 by converting the data type of the weights of the third neural network model 43. For example, if the weights of the third neural network model 43 are 32-bit floating-point values, the neural network model compression module 233 can convert the weights of the third neural network model 43 into 8-bit integer values. That is, the neural network model compression module 233 can obtain the second neural network model 42 by compressing the size of the weights of the third neural network model 43. In this way, the neural network model compression module 233 can obtain the second neural network model 42, which is a compressed form of the third neural network model 43. Then, the processor 230 can send the second neural network model 42 to the electronic device 100 through the communication interface 210. Therefore, even when the memory 140 of the electronic device 100 has limited capacity, the electronic device 100 can update the first neural network model 41 to the second neural network model 42.

[0084] Figure 4A This diagram illustrates a method for recognizing and learning images 15 according to an embodiment of the present disclosure. By inputting multiple captured images 10 into a first neural network model 41, the electronic device 100 can obtain classification values ​​and probability values ​​for objects included in multiple corresponding captured images 10. For example, by inputting the first captured image 10-1 into the first neural network model 41, the electronic device 100 can obtain a classification value ("clothes") and a probability value ("0.1") for the first object ob1. By inputting the second captured image 10-2 into the first neural network model 41, the electronic device 100 can obtain a classification value ("wire") and a probability value ("0.2") for the second object ob2. By inputting the third captured image 10-3 into the first neural network model 41, the electronic device 100 can obtain a classification value ("polyethylene") and a probability value ("0.3") for the third object ob3. By inputting the fourth captured image 10-4 into the first neural network model 41, the electronic device 100 can obtain a classification value ("clothes") and a probability value ("0.9") for the fourth object ob4. Meanwhile, in Figure 4A In this context, for ease of explanation, the classification value for each object is expressed using words that indicate that object; however, as mentioned above, the classification value can be a numerical value corresponding to each object.

[0085] Then, the electronic device 100 can identify the learning image 15 based on the probability values ​​of the objects included in the plurality of corresponding captured images 10. Specifically, the electronic device 100 can identify an image in the plurality of captured images 10 that includes objects with probability values ​​less than a predetermined value as the learning image 15. Here, the predetermined value can represent a value used to evaluate the recognition rate or learning degree of the electronic device 100 for objects. Additional training is required for the electronic device 100 for objects with probability values ​​smaller than the predetermined value. For example, with a predetermined value of 0.6, the electronic device 100 can identify the first captured image 10-1, the second captured image 10-2, and the third captured image 10-3, which respectively include the first object ob1, the second object ob2, and the third object ob3 with probability values ​​less than 0.6, as the learning image 15. In contrast, the electronic device 100 may not identify the fourth captured image 10-4, which includes the fourth object ob4 with a probability value greater than 0.6, as the learning image 15.

[0086] Figure 4B and Figure 4C This is a diagram illustrating a method for obtaining learning data of an electronic device according to an embodiment of the present disclosure.

[0087] refer to Figure 4B The electronic device 100 can cluster feature values ​​by mapping feature values ​​of multiple corresponding learning images 15 (i.e., learning data 15-1, 15-2, and 15-3) to vector space 410. Specifically, the electronic device 100 can cluster feature values ​​by mapping first to third feature values ​​V1, V2, and V3 corresponding to objects ob1, ob2, and ob3, respectively, to vector space 410. Therefore, multiple clusters C1, C2, and C3, each including at least one feature value, can be formed in vector space 410. Here, as described above, the first to third feature values ​​V1, V2, and V3 can be obtained by using a first neural network model 41. Note that although in Figure 4B The vector space 410 is shown as a three-dimensional space, but this is only an example, and the vector space 410 can be a space with dimensions greater than or equal to three. Furthermore, each axis x, y, z of the vector space 410 can represent various factors, such as factors like color, size, texture, and shape.

[0088] refer to Figure 4C The electronic device 100 can identify clusters with a cohesion greater than a predetermined value among multiple clusters C1, C2, and C3 existing in the vector space 410. For example, the electronic device 100 can identify the first cluster C1 and the second cluster C2 among multiple clusters C1, C2, and C3.

[0089] The electronic device 100 can acquire captured images corresponding to the feature values ​​included in the first cluster C1 and the second cluster C2 as learning data 30. Here, the electronic device 100 can identify the feature value that is closest to the average value of the feature values ​​included in the first cluster C1. For example, the electronic device 100 can identify the 1-1 feature value V included in the first cluster C1. 1-1 Then, electronic device 100 can obtain features including 1-1 characteristic value V. 1-1 The corresponding learning image 15 of the object is used as the first learning data 30-1. Here, the first learning data 30-1 may include information about the 1-1 feature value V. 1-1 The location information of region R1 of the object in the cluster. Similarly, the electronic device 100 can identify the 2-1 feature value V that is closest to the average of the feature values ​​included in the second cluster C2. 2-1 And obtain eigenvalues ​​V including 2-1 eigenvalues. 2-1 The corresponding learning image 15 is used as the second learning data 30-2.

[0090] At the same time, the electronic device 100 can identify multiple feature values ​​V for each cluster. n And obtain the corresponding identified feature values ​​V n The corresponding multiple learning images 15 serve as learning data 30. Here, the electronic device 100 can identify images present at a distance of from each cluster C. n Feature value V within a predetermined distance from the center n And obtain the feature value V that is identified. n Multiple learning images 15 of the corresponding objects are used as learning data 30.

[0091] Figure 5A This is a diagram illustrating a method for recognizing and learning images according to another embodiment of the present disclosure. Specifically, Figure 5A This is a diagram illustrating a method for recognizing and learning images when multiple objects are included in the captured image 10.

[0092] If the captured image 10 includes multiple objects, and if one of the objects has a probability value less than a predetermined value, the electronic device 100 can identify the captured image 10 as the learning image 15. (Reference) Figure 5A The captured image 10 may include a fifth object ob5, a sixth object ob6, and a seventh object ob7. Furthermore, the probability value of the fifth object ob5 may be 0.2, the probability value of the sixth object ob6 may be 0.7, and the probability value of the seventh object ob7 may be 0.8. In the example where the predetermined probability value is 0.7, since the probability value of the fifth object ob5 is less than the predetermined value, the electronic device 100 can recognize the captured image 10 as the learning image 15.

[0093] Figure 5B This is a diagram illustrating a clustering method according to another embodiment of the present disclosure. Electronic device 100 can cluster feature values ​​by mapping feature values ​​of objects included in a learned image to a vector space. Here, electronic device 100 can identify feature values ​​V of objects with probability values ​​less than a predetermined value. n The identified feature values ​​are then mapped to vector space 510. (Reference) Figure 5B The electronic device 100 can identify the fifth feature value V5 of the fifth object ob5, which has a probability value smaller than a predetermined value, and cluster the feature value by mapping the fifth feature value V5 to the vector space 510. Conversely, the electronic device 100 may not map the sixth feature value V6 and the seventh feature value V7 of the sixth object ob6 and the seventh object ob7 to the vector space 510 because these feature values ​​have a probability value larger than the predetermined value.

[0094] As described above, the electronic device 100 can acquire learning data 30 and send it to the external device 200. The external device 200 can then obtain a second neural network model 42 based on the learning data 30. The method by which the external device 200 obtains the second neural network model 42 will be described below.

[0095] Figure 6A This is a diagram illustrating a method for obtaining labeled data according to an embodiment of the present disclosure.

[0096] External device 200 can obtain learning data 30, including first learning data 30-1 and second learning data 30-2, from electronic device 100. External device 200 can obtain classification values ​​and probability values ​​of the objects included in learning data 30 by inputting learning data 30 into fourth neural network model 44. For example, external device 200 can obtain the classification value ("clothes") and probability value ("0.8") of a first object ob1 by inputting first learning data 30-1 into fourth neural network model 44, and external device 200 can obtain the classification value ("wire") and probability value ("0.5") of a second object ob2 by inputting second learning data 30-2 into fourth neural network model 44. As described above, fourth neural network model 44 is a model with higher performance than first neural network model 41, and fourth neural network model 44 can include more layers than first neural network model 41.

[0097] External device 200 can identify objects with a probability value greater than a predetermined value. For example, if the predetermined value is 0.7, external device 200 can identify a first object ob1 with a probability value greater than 0.7. Then, external device 200 can perform a labeling operation on the identified object. Here, external device 200 can match first learning data 30-1 including the first object ob1 and the classification value ("clothes") of the first object ob1, and store them. Therefore, external device 200 can obtain labeled data 31.

[0098] refer to Figure 6B External device 200 can train a third neural network model 43 based on labeled data 31. The third neural network model 43 can be trained to recognize objects included in the labeled data 31. Here, external device 200 can train the third neural network model 43 based on supervised learning.

[0099] Simultaneously, the external device 200 can train a third neural network model 43 based on the labeled data 31 and the learning data 32. Here, the learning data 32 is the data used to train the first neural network model 41, and the learning data 32 can be pre-stored in the external device 200. Therefore, it is possible to prevent the third neural network model 43 from being biased or overfitting relative to the labeled data 31.

[0100] External device 200 can obtain second neural network model 42 by compressing the weights of third neural network model 43. Specifically, external device 200 can obtain weights of second neural network model 42 by reducing the size of the data type of weights (or parameters) of third neural network model 43. (See reference) Figure 6C The third neural network model 43 can have 1-1 weighted values ​​w1-1 and 2-1 weighted values ​​w2-1 of the third data type 73. For example, the third data type 73 can be a 32-bit floating-point type. Furthermore, the second neural network model 42 can have 1-2 weighted values ​​w1-2 and 2-2 weighted values ​​w2-2 of the second data type 72. For example, the second data type 72 can be an 8-bit integer type.

[0101] like Figure 6CAs shown, when external device 200 converts third data type 73 to second data type 72, it can obtain 1-2 weighted value w1-2 and 2-2 weighted value w2-2 from 1-1 weighted value w1-1 and 2-1 weighted value w2-1, respectively. Here, external device 200 can obtain the values ​​closest to 1-1 weighted value w1-1 and 2-1 weighted value w2-1 as 1-2 weighted value w1-2 and 2-2 weighted value w2-2, respectively. For example, external device 200 can obtain the value of second data type 72 that is closest to 1-1 weighted value w1-1 as 1-2 weighted value w1-2. Here, 1-1 weighted value w1-1 and 1-2 weighted value w1-2 can have the same value. Furthermore, external device 200 can obtain 2-2 weighted value w2-2, which is the value of second data type 72 that is closest to 2-1 weighted value w2-1.

[0102] External device 200 can send information about the obtained second neural network model 42 to electronic device 100. Specifically, external device 200 can send the second neural network model 42 to electronic device 100. Alternatively, external device 200 can send the parameters (or weighting values) of the second neural network model 42 to electronic device 100. Simultaneously, external device 200 can send the second neural network model 42 or its parameters (or weighting values) to electronic device 100 based on the available capacity of memory 140. For example, if the available capacity of memory 140 is less than a predetermined value, external device 200 can send only parameter information other than the configuration information of the second neural network model 42 to electronic device 100. Alternatively, if the available capacity of memory 140 is greater than a predetermined value, external device 200 can send the second neural network model 42 itself, including both its configuration information and parameter information.

[0103] The electronic device 100 can update the first neural network model 41 based on received information about the second neural network model 42. For example, if the second neural network model 42 is received, the electronic device 100 can store the second neural network model 42 in the memory 140 and delete the pre-stored first neural network model 41. Alternatively, upon receiving parameters of the second neural network model 42, the electronic device 100 can update the first neural network model 41 by changing the parameters of the first neural network model 41 to the parameters of the second neural network model 42.

[0104] The above describes the operation of electronic and external devices.

[0105] The following section describes methods for controlling electronic and external devices.

[0106] Figure 7AThis is a flowchart illustrating a control method for an electronic device according to an embodiment of the present disclosure. (Reference) Figure 7A During operation S710, the electronic device 100 can acquire multiple captured images. Here, the electronic device 100 can acquire captured images 10 showing the surroundings of the electronic device 100 via the camera 110. Alternatively, the electronic device 100 can acquire multiple captured images 10 from an external device 200 via the communication interface 120.

[0107] In operation S720, the electronic device 100 can obtain feature values ​​V of the object ob included in the multiple captured images 10 by inputting the multiple captured images 10 into the first neural network model 41. n The classification value and the probability value of the classification value are used to classify the object ob. Here, when the feature value V of object ob... n When input into the output layer of the first neural network model 41, a classification value (or predicted category) and a probability value can be obtained. Therefore, the classification value and the probability value can correspond to the feature value V, respectively. n .

[0108] In operation S730, the electronic device 100 can identify multiple learning images 15 from multiple captured images 10 based on the obtained probability values. Here, the electronic device 100 can identify the learning images 15 as captured images 10 that include an object ob with a probability value smaller than a predetermined value.

[0109] In operation S740, electronic device 100 can identify cluster feature values ​​by mapping feature values ​​of objects included in multiple correspondingly identified learning images 15 to a vector space (e.g., vector space 410). Here, electronic device 100 can identify cluster C among multiple clusters existing in the vector space that has a cohesion greater than a predetermined value. n And identify in the identified cluster C n The eigenvalues ​​included in it.

[0110] In operation S750, electronic device 100 can obtain learning data 30 from multiple learning images 15 based on the identified feature values. Here, electronic device 100 can identify clusters C. n The feature value that is closest to the average of the multiple feature values ​​included in the data is then selected. The electronic device 100 can then obtain an image of the object to be learned, including the object corresponding to the identified feature value, as learning data 30.

[0111] In operation S760, the electronic device 100 can send the acquired learning data 30 to an external device. Here, the learning data 30 may include the raw data of the image to be learned (or the captured image) and the positional information of objects within the image to be learned.

[0112] In operation S770, the electronic device 100 can receive information about the second neural network model 42 from the external device 200. Then, in operation S780, the electronic device 100 can update the first neural network model 41 based on the information about the second neural network model 42. Here, upon receiving the second neural network model 42, the electronic device 100 can store the second neural network model 42 in the memory 140 and delete the pre-stored first neural network model 41. Alternatively, upon receiving the parameters of the second neural network model 42, the electronic device 100 can update the first neural network model 41 by changing the parameters of the first neural network model 41 to the parameters of the second neural network model 42.

[0113] Figure 7B This is a flowchart illustrating a control method for an electronic device according to an embodiment of the present disclosure. (Reference) Figure 7B The electronic device 100 can recognize multiple learning images 15 in operation S730 and store the recognized learning images 15 in operation S735. Here, the electronic device 100 can store multiple recognized learning images 15 in memory 140.

[0114] Then, in operation S737, electronic device 100 can identify whether idle resources exist. Specifically, electronic device 100 can identify whether the idle value of processor 150 is greater than or equal to a predetermined value (e.g., 90%). Alternatively, electronic device 100 can identify whether the usage of memory 140 is less than or equal to a predetermined value (e.g., 10%).

[0115] If no free resource is detected, electronic device 100 can return. Figure 7A In operation S710, multiple captured images 10 are acquired. Then, in operation S730, the electronic device 100 can identify multiple learning images 15 from the multiple acquired captured images 10, and store these images 10 in operation S735.

[0116] If the existence of idle resources is identified, the electronic device 100 can identify clustered feature values ​​V by mapping the feature values ​​of objects included in multiple corresponding learning images 15 to a vector space (e.g., vector space 410). n ,like Figure 7A In operation S740, the electronic device 100 then obtains learning data 30 from multiple learning images 15 based on the identified feature values ​​in operation S750. Then, in operation S755, the electronic device 100 stores the obtained learning data 30.

[0117] In operation S757, electronic device 100 can identify whether it is in standby mode. Here, standby mode can be a mode where electronic device 100 is being charged by an external power source. If electronic device 100 is not identified as being in standby mode, it can return to... Figure 7A In operation S710, multiple captured images 10 are acquired. Then, in operation S730, the electronic device 100 can identify multiple learning images 15 from the multiple acquired captured images 10, and store these images in operation S735.

[0118] If electronic device 100 is identified as being in standby mode, electronic device 100 can send the acquired learning data 30 to external device 200, such as... Figure 7A Operation S760. Here, the acquired learning data 30 is the learning data 30 stored in operation S755. Then, in operation S770, the electronic device 100 can receive information about the second neural network model 42 from the external device 200, and in operation S780, update the first neural network model 41 based on the information about the second neural network model 42. Meanwhile, as described above, when the electronic device 100 enters standby mode, the electronic device 100 sends the learning data 30 to the external device 200. However, this is only an example; when a user instruction for updating the first neural network model 41 is entered, the electronic device 100 can send the stored learning data 30 to the external device 200.

[0119] Figure 8 This is a sequence diagram illustrating a neural network model update system according to an embodiment of the present disclosure. In operation S810, the electronic device 100 can acquire multiple captured images 10. Since this operation corresponds to operation S710, a detailed explanation thereof will be omitted.

[0120] In operation S820, the electronic device 100 can obtain learning data 30 from multiple captured images 10 by using a first neural network model 41. Specifically, the electronic device 100 can... Figure 7A Operations S720, S730, S740, and S750 obtain learning data 30. Then, in operation S830, electronic device 100 can send the learning data 30 to external device 200. In operation S840, external device 200 can obtain information about the second neural network model 42 based on the learning data 30. Here, external device 200 can obtain labeled data 31 from the learning data 30 using a fourth neural network model 44. Then, external device 200 can train a third neural network model 43 based on the labeled data 31, compress the weights of the third neural network model 43, and obtain information about the second neural network model 42.

[0121] Then, in operation S850, electronic device 100 can receive information about the second neural network model 42 from external device 200, and in operation S860, update the first neural network model 41 based on the information about the second neural network model 42.

[0122] Furthermore, the various embodiments described above can be implemented using software, hardware, or a combination thereof, on a recording medium readable by a computer or similar device. In some cases, the embodiments described herein can be implemented as the processor itself. According to software implementation, the embodiments such as processes and functions described herein can be implemented as separate software modules. Each software module can perform one or more functions and operations described herein.

[0123] Furthermore, according to the various embodiments of this disclosure described above, computer instructions for performing processing operations can be stored in a non-transitory computer-readable medium. When the computer instructions stored in such a non-transitory computer-readable medium are executed by a processor of a particular machine, the instructions can cause the particular machine to perform the processing operations according to the various embodiments described above.

[0124] Non-transitory computer-readable media refers to media that store data semi-permanently and can be read by a machine, but does not refer to media that store data for short periods, such as registers, caches, and memory. Specific examples of non-transitory computer-readable media include CDs, DVDs, hard drives, Blu-ray discs, USB drives, memory cards, and ROMs.

[0125] Simultaneously, machine-readable storage media can be provided in the form of non-transitory storage media. Here, the term "non-transitory storage media" refers only to tangible devices and does not include signals (e.g., radio waves), and the term does not distinguish between cases where data is stored semi-permanently in the storage medium and cases where data is temporarily stored. For example, "non-transitory storage media" may include buffers in which data is temporarily stored.

[0126] Furthermore, according to embodiments, methods according to various embodiments disclosed herein can be provided while being included in a computer program product. A computer program product refers to a product that can be traded between a seller and a buyer. A computer program product may be distributed in the form of a machine-readable storage medium (e.g., a compact disk read-only memory (CD-ROM)) or through an app store (e.g., the Play Store). TMThis can be done directly online (e.g., by downloading or uploading) or between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., a downloadable application) may be stored at least temporarily in a machine-readable storage medium (such as the manufacturer's server, app store server, and the memory of a relay server), or may be temporarily generated.

[0127] Furthermore, while embodiments of this disclosure have been shown and described, this disclosure is not limited to the specific embodiments described above, and it is clear that those skilled in the art to which this disclosure pertains can make various modifications to this disclosure without departing from the spirit of the disclosure as described in the appended claims. Moreover, such modifications should not be interpreted independently of the technical concept or prospects of this disclosure.

Claims

1. A method for controlling an electronic device, the method comprising: Obtain multiple images, wherein the multiple images include one or more objects; Multiple images are input into a first neural network model for object recognition, and based on the input of multiple images into the first neural network model for object recognition, feature values ​​of each of one or more objects, predicted categories of each of one or more objects based on the corresponding obtained feature values, and probability values ​​of predicted categories of each of one or more objects are obtained. Identify one or more learned images from multiple images based on the obtained probability values; One or more clusters of feature values ​​are identified by mapping the feature values ​​of one or more objects included in one or more identified learning images to a vector space; Learning data is obtained from one or more identified learning images based on the obtained feature values; Send the acquired learning data to an external device; Receive information about the second neural network model from an external device; and The first neural network model is updated based on the received information about the second neural network model.

2. The method according to claim 1, in, Recognizing one or more learned images also includes: Identify one or more learned images, including one or more objects whose probability value is less than a predetermined value.

3. The method according to claim 1, in, The acquisition of learning data also includes: Identify at least one cluster with a cohesion greater than a predetermined value among one or more clusters of feature values ​​including feature values ​​of one or more objects included in one or more identified learning images; and One or more images corresponding to the feature values ​​included in at least one cluster of the identified feature values ​​are obtained as learning data.

4. The method according to claim 3, in, Obtaining one or more images corresponding to the feature values ​​included in at least one cluster of the identified feature values ​​as learning data includes: The image corresponding to the average closest feature value included in at least one cluster of the identified feature values ​​is identified and obtained as learning data.

5. The method according to claim 3, further comprising: Store the acquired learning data. The storage includes: In at least one cluster of identified feature values, a cluster of feature values ​​including a predetermined number of feature values ​​is identified, and an image corresponding to the feature values ​​included in the identified cluster is stored.

6. The method according to claim 1, in, The learning data is at least one of one or more learning images, and wherein the learning data includes positional information of pixels corresponding to objects included in the at least one image.

7. The method according to claim 1, in, The second neural network model is a model obtained based on a third neural network model that has higher accuracy than the first neural network model, wherein the third neural network model is trained based on the obtained learning data.

8. An electronic device, comprising: A communication interface, including at least one circuit; Memory, including at least one instruction; as well as processor, The processor is configured as follows: Obtain multiple images, wherein the multiple images include one or more objects; By inputting multiple images into a first neural network model for object recognition, feature values ​​of each of one or more objects, predicted categories of each of one or more objects based on the corresponding obtained feature values, and probability values ​​of predicted categories of each of one or more objects are obtained. Identify one or more learned images from multiple images based on the obtained probability values; One or more clusters of feature values ​​are identified by mapping the feature values ​​of one or more objects included in one or more identified learning images to a vector space; Learning data is obtained from one or more identified learning images based on the obtained feature values; The acquired learning data is sent to external devices via a communication interface; Receive information about the second neural network model from an external device; and The first neural network model is updated based on the received information about the second neural network model.

9. The electronic device according to claim 8, in, The processor is also configured to: Identify one or more learned images, including one or more objects whose probability value is less than a predetermined value.

10. The electronic device according to claim 8, in, The processor is also configured to: Identify at least one cluster with a cohesion greater than a predetermined value among one or more clusters of feature values ​​including feature values ​​of one or more objects included in the identified learning image; and One or more images corresponding to the feature values ​​included in at least one cluster of the identified feature values ​​are obtained as learning data.

11. The electronic device according to claim 10, in, The processor is also configured to: The image corresponding to the average closest feature value included in at least one cluster of the identified feature values ​​is identified and obtained as learning data.

12. The electronic device according to claim 10, in, The processor is also configured to: Identify clusters of feature values ​​that include a predetermined number of feature values ​​among at least one identified cluster, and store images corresponding to the feature values ​​included in the identified clusters of feature values.

13. The electronic device according to claim 9, in, The learning data is at least one of one or more learning images, and wherein the learning data includes positional information of pixels corresponding to objects included in the at least one image.

14. The electronic device according to claim 9, in, The second neural network model is a model obtained based on a third neural network model that has higher accuracy than the first neural network model, wherein the third neural network model is trained based on the obtained learning data.

15. A neural network model update system including electronic devices and external devices, in, The electronic device is configured to: Multiple images are acquired, and learning data is obtained from these images using a first neural network model. This learning data is then sent to an external device, from which information about a second neural network model is received. Based on this information, the first neural network model is updated. The external device is configured as follows: The learning data is input into the fourth neural network model to obtain labeled data. The third neural network model is trained based on the labeled data. Information about the second neural network model is obtained by compressing the third neural network model. The information about the second neural network model is then sent to an electronic device.

Images