Apparatus and method for learning image recognition

Abstract

The invention relates to an apparatus and method for learning image recognition. According to the present invention, it is possible to analyze the cause of degradation of recognition performance found in an AI-based learning and verification process to extract features of an image analyzed by the AI, thereby regenerating an image reflecting the extracted features, and to perform the learning and verification process, thereby improving image recognition performance.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of priority from Korean Patent Application No. 10-2023-0136666 filed in the Korean Intellectual Property Office on October 13, 2023, the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present disclosure relates to a technology for learning image recognition, and more specifically, to an apparatus and method for learning image recognition, which is capable of improving the recognition rate of each element to be detected in an autonomous driving vehicle through an artificial intelligence (AI)-based image recognition model (AI-based image recognition model). Background Art

[0004] In order to solve complex problems or solve unknown problems, various studies and research have been conducted on realizing the human way of thinking through computing devices. One of them is the technology of neural networks that are modeled to realize the behavior of human neurons in hardware. In order to approach the learning ability of humans, the technology of neural networks is used to model a learning model based on a neural network that is obtained by mathematically expressing the synaptic neurotransmission process. The neural network can map input and output to each other through learning.

[0005] The most representative thinking behavior in a neural network is the recognition behavior of identifying an object. This kind of neural network is called a recognition model. The recognition model is configured and machine trained to recognize what an object is or what the behavior of an object is.

[0006] Image recognition technology is a core technology in autonomous driving. Image recognition algorithms based on deep neural networks (DNN) and image data obtained on real roads are used to train / verify image recognition algorithms. For example, image recognition algorithms can be trained and verified through post-processing operations and labeling operations on image data obtained on real roads. Recently, development methods using virtual simulations have been used to efficiently perform training and verification processes.

[0007] However, when verifying simulated synthetic images, the scope of verifying simulated synthetic images is limited due to the quality difference between real images and synthetic images. In addition, the cost of generating a database (DB) of real images is excessively required, and some data is disposable. Therefore, when developing an image recognition algorithm applied to a new car, it is continuously incurred to generate a DB of real images for training and verification. Summary of the invention

[0008] The present disclosure has been made to solve the above-mentioned problems occurring in the related art while keeping the advantages achieved by the related art intact.

[0009] One aspect of the present disclosure provides a device and method for learning image recognition, which can improve the recognition rate of elements to be detected in an autonomous driving vehicle through an artificial intelligence (AI)-based image recognition model (hereinafter, an AI-based image recognition model).

[0010] Another aspect of the present disclosure provides an apparatus and method for learning image recognition, which is capable of analyzing the causes of degradation of recognition performance discovered during an AI-based learning and verification process to extract features of an image analyzed by AI to regenerate an image reflecting the extracted features, and performing a learning and verification process to improve image recognition performance.

[0011] Another aspect of the present disclosure provides an apparatus and method for learning image recognition, which is capable of training an image recognition model by reflecting the latent vector of each element that is not recognized or is incorrectly recognized when generating an enhanced image for a synthetic image generated by simulation, thereby quantitatively / qualitatively improving the synthetic image, thereby improving the performance of the image recognition model.

[0012] The technical problems to be solved by the present disclosure are not limited to the above-mentioned problems, and any other technical problems not mentioned herein will be clearly understood by those skilled in the art to which the present disclosure belongs from the following description.

[0013] According to one aspect of the present disclosure, a device for learning image recognition includes: a conversion device for generating an enhanced image that enhances specific elements by converting an input image into an enhanced image via a generative artificial intelligence (AI) model to enhance a latent vector of at least one element, a training device for training the image recognition model based on AI by utilizing the enhanced image; a determination device for comparing a recognition rate of at least one element in the enhanced image provided by the image recognition model and a target recognition rate preset for the at least one element, and an analysis device for analyzing the latent vector of at least one element with a recognition rate lower than the target recognition rate when at least one element with a recognition rate lower than the target recognition rate exists in the enhanced image, and providing the analyzed latent vector of the at least one element to the conversion device so that the latent vector of the at least one element is reflected in the generative AI model.

[0014] According to an embodiment, the conversion device can also generate a new enhanced image by reflecting the latent vector of at least one element analyzed by the analysis device in the generative AI model, and convert the input image into the new enhanced image that enhances the latent vector of at least one element analyzed.

[0015] According to an embodiment, the determination device may further compare the recognition rate of at least one element in the new enhanced image provided by the image recognition model with the target recognition rate. When at least one element with a recognition rate lower than the target recognition rate exists in the new enhanced image, the analysis device may further analyze the latent vector of at least one element with a recognition rate lower than the target recognition rate, and may provide the analyzed latent vector of at least one element to the conversion device.

[0016] According to an embodiment, the analyzing device may further iteratively provide the latent vector of the at least one element being analyzed to the converting device until the recognition rate of the at least one element is equal to or greater than the target recognition rate of the at least one element.

[0017] According to an embodiment, the target recognition rate may be set differently according to the autonomous driving level and for at least one element, and the input image may include a simulated synthetic image.

[0018] According to another aspect of the present disclosure, a method for learning image recognition includes: a conversion device converting an input image into an enhanced image via a generative artificial intelligence (AI) model to generate an enhanced image that enhances a specific element to enhance a latent vector of at least one element; a training device training an image recognition model based on the AI ​​by utilizing the enhanced image; a determination device comparing a recognition rate of the at least one element provided by the image recognition model with a target recognition rate preset for the at least one element; and, in response to the presence of the at least one element having a recognition rate lower than the target recognition rate in the enhanced image, an analysis device analyzing the latent vector of the at least one element; and the conversion device reflecting the latent vector of the at least one element in the generative AI model.

[0019] According to an embodiment, generating an enhanced image may include: generating a new enhanced image by a conversion device by reflecting a latent vector of at least one element being analyzed in a generative AI model; and converting the input image into the new enhanced image that enhances the latent vector of the at least one element being analyzed by the conversion device.

[0020] According to an embodiment, the comparison may include: a determination device compares a recognition rate of at least one element in a new enhanced image provided by an image recognition model with a target recognition rate, and the analysis may include: in response to the presence of at least one element having a recognition rate lower than the target recognition rate in the new enhanced image, an analysis device analyzes a potential vector of the at least one element.

[0021] According to an embodiment, the generating and reflecting may be iteratively performed until the recognition rate of at least one element is equal to or greater than the target recognition rate of the at least one element.

[0022] According to an embodiment, the target recognition rate may be set differently according to the autonomous driving level and for at least one element, and the input image may include a simulated synthetic image. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other objects, features and advantages of the present disclosure will become more apparent from the following detailed description in conjunction with the accompanying drawings:

[0024] Figure 1 is a diagram showing explored domains and unexplored domains according to autonomous driving levels;

[0025] Figure 2 is a block diagram showing an apparatus for learning image recognition according to an embodiment of the present disclosure;

[0026] Figure 3 is a flowchart illustrating the operation of a method for learning image recognition according to another embodiment of the present disclosure;

[0027] Figure 4 is a diagram showing a method according to an embodiment Figure 3 A flowchart of the operation in S350;

[0028] Figure 5A and Figure 5B are a diagram showing a target recognition rate of each element and a diagram showing a recognition result, respectively;

[0029] Figure 6 is a diagram illustrating a process of analyzing a latent vector of an element; and

[0030] Figure 7 is a block diagram of a computing system executing a method for learning image recognition according to another embodiment of the present disclosure. DETAILED DESCRIPTION

[0031] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings to allow those skilled in the art to easily reproduce the present disclosure. However, the present disclosure may be implemented in various forms and is not limited to the embodiments described herein.

[0032] In the following description of the present disclosure, when it is determined that the detailed description of the related known configuration or function may make the subject matter of the present disclosure unclear, its details may be omitted. In addition, in order to clearly describe the present disclosure, parts irrelevant to the present disclosure are omitted, and similar reference numerals are assigned to similar parts throughout the entire specification.

[0033] It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element or there may be intervening elements. In addition, a component "including" or "having" another component may mean that the component further includes another component without excluding the other component, unless otherwise specified.

[0034] In the present disclosure, unless otherwise defined, the term "first" or "second" is used to distinguish one component from another component without limiting the order or importance between the components. Therefore, within the scope of the present disclosure, the "first component" according to an embodiment may be referred to as the "second component" according to another embodiment, and the "second component" according to an embodiment may be referred to as the "first component" according to another embodiment.

[0035] According to the present disclosure, components or devices different from each other are provided to clearly describe the features of each component, and it does not mean that the components or devices are separated from each other. In other words, multiple components or devices are integrated and implemented in a hardware or software unit. Alternatively, a component or device is split and implemented in units of multiple hardware or software. Therefore, unless otherwise specified, even embodiments with components or devices arranged integrally or separately from each other fall within the scope of the present disclosure. Each device can be configured to perform specific operations or actions by installing software, firmware, hardware (e.g., processor) or a combination thereof in the device, and the software, firmware, hardware (e.g., processor) or a combination thereof causes the device to perform actions in operation.

[0036] In the present disclosure, the components described according to various embodiments do not refer to necessary components, and some components may be provided selectively. Therefore, embodiments including a subset of the components described according to the embodiments are included in the present disclosure. In addition, embodiments including components added to the components described according to the various embodiments are also included in the scope of the present disclosure.

[0037] In the present disclosure, the positional relationship (eg, top, bottom, left, or right) indicated in the specification is provided for illustrative purposes only. When the drawings of the present disclosure are shown in reverse, the positional relationship described in the present disclosure can be interpreted in reverse.

[0038] In the present disclosure disclosed herein, each of the expressions "A or B", "at least one of A and B", "at least one of A or B", "A, B or C", "at least one of A, B and C", or "at least one of A, B or C" etc. used herein may include any one or all combinations of the items arranged together in the relevant expressions of the expressions.

[0039] The technology of image recognition, which is a core technology in autonomous driving, tends to utilize AI technology represented by deep neural network (DNN) technology. The technology of image recognition requires images with a large amount of big data provided in a different manner. Recently, the technology of image recognition adopts supervised learning technology that generates an image data set called "real image" and trains a recognition algorithm through labeling processing and pre / post processing of data required for object recognition.

[0040] However, the technology of recognizing images based on AI requires a large amount of cost, such as data acquisition cost, cost for pre-processing / post-processing, or labeling cost to generate training / validation data. In addition, when the required performance is not met in training and validation, the performance should be enhanced through an additional training process. However, at present, there is no suggestion to analyze the reasons for improving the recognition performance and generate an image reflecting the information for enhancing the analyzed reasons, so that a retraining process / revalidation process is performed.

[0041] In addition, currently, for the embodiments that are not recognized and / or misrecognized, a simple approximation method is used to additionally train a larger number of images to improve the recognition performance. As the level of autonomous driving increases, the complexity and difficulty of developing technology increases, and the number of evaluation items to be verified increases exponentially. For example, based on Figure 1 As the autonomous driving level shown is increased as the ADAS level is shown in the explored domain / unexplored domain, the unexplored domain that makes image recognition algorithms (which can be processed by humans) difficult increases. Therefore, it is difficult to ensure the recognition performance quality required for each autonomous driving level.

[0042] Furthermore, it is actually difficult to train / verify the entire domain of recognition algorithms based on AI by verifying real cars based on human capabilities. Therefore, training methods and evaluation methods through virtual simulations for effective and efficient development have been proposed and utilized. In addition, development methods using synthetic images generated through simulations have been performed. However, due to the quality difference between synthetic images and real images, the scope of using synthetic images is limited.

[0043] According to the subject matter of an embodiment of the present disclosure, the causes of degradation of recognition performance discovered during an AI-based learning and verification process can be analyzed to extract features of an image analyzed by AI, thereby regenerating an image reflecting the extracted features, and performing a learning and verification process to improve image recognition performance.

[0044] In addition, according to an embodiment of the present disclosure, the image recognition model can be trained by reflecting the features (latent vectors) of each element that is not recognized or incorrectly recognized when an enhanced image is generated for a synthetic image generated by simulation, thereby quantitatively / qualitatively improving the achieved synthetic image, thereby improving the performance of the image recognition model.

[0045] In addition, according to an embodiment of the present disclosure, the reason why the image database (DB) fails to meet the recognition rate can be analyzed. Therefore, the recognition rate of each element can be improved by an image recognition algorithm (or model). In addition, since a database with substantially real images is generated, the replacement rate of real images can be increased, and thus the cost and time for generating images can be reduced.

[0046] Will refer to Figures 2 to 6 An apparatus and method for learning image recognition according to an embodiment of the present disclosure is described.

[0047] Figure 2 is a block diagram illustrating an apparatus for learning image recognition according to an embodiment of the present disclosure.

[0048] refer to Figure 2 According to an embodiment of the present disclosure, an apparatus 200 for learning image recognition includes a conversion device 210 , a learning device 220 , a determination device 230 , an analysis device 240 , and a storage device 250 .

[0049] The storage device 250 (which is a device for storing data related to the device of the present disclosure) stores all data related to the technology of the present disclosure, wherein all data include a generative AI model, an AI-based image recognition model (or algorithm), a feature information database (DB), an enhanced image generated by a generative model, a synthetic image input as an input image, (such as a vehicle, a pedestrian, or a raining state, about elements to be recognized in the image) information, or (such as a guardrail shadow, a building part, or a tunnel part, about driving conditions) information. The storage device 250 can store all data related to the technology of the present disclosure or required to execute the technology of the present disclosure as well as the above data.

[0050] The conversion device 210 converts an input image such as a simulated synthetic image input to the device 200 for learning image recognition into an enhanced image with enhanced specific elements through a generative AI model, for enhancing a preset latent vector of each element.

[0051] According to an embodiment, for the converted enhanced image, the conversion device 210 may receive the latent vector of the element analyzed by the analysis device 240 from the analysis device 240, and reflect the latent vector in the generative AI model, thereby generating a new enhanced image (e.g., a new enhanced image in which the latent vector of the element analyzed by the analysis device 240 is enhanced) by converting the input image into another enhanced image. In addition, when the image recognition algorithm is re-evaluated by the new enhanced image, and therefore the recognition rate of the new enhanced image is not satisfied, the latent vector of the element analyzed for the new enhanced image is received from the analysis device 240. In this case, the conversion device 210 may reflect the latent vector of the received element in the generative AI model, thereby regenerating a new enhanced image for the input image. The above process is iteratively performed until the recognition rate of the element included in the image is higher than the preset target recognition rate.

[0052] The generative AI model is a conversion model having a function of realizing a synthetic image, and is a generative model based on a latent vector, and is used to make a customized image to enhance elements showing recognition weaknesses in the image recognition model, so that the recognition performance is improved. According to the generative AI model, new data that does not exist in the existing learning DB should be created based on the latent vector of the element extracted by the analysis device 240, so as to optimally train the image recognition model, and different images beyond the scope of the specification or human thinking ability can be generated.

[0053] The training device 220 can train or retrain the image recognition model based on AI by using an enhanced image, which is obtained by the conversion of the conversion device 210, for example, by using a realized synthetic image in which the latent vector of the element is reflected. In addition, the training device 220 can provide or output a recognition result (for example, a recognition rate of each element) for each element to be detected included in the enhanced image by using an AI-based image recognition model. In other words, when the performance of the image recognition model is insufficient, the training device 220 additionally receives a customized image (i.e., an enhanced image) from the conversion device 210 for improving the recognition performance, and performs supplementary training on the image recognition model, thereby iteratively supplementing the image recognition model so that the performance of the image recognition model is improved.

[0054] According to an embodiment, the training device 220 may be a component that not only uses an enhanced image to train an AI-based image recognition model but also verifies the trained AI-based image recognition model. For example, the training device 220 may train or retrain an AI-based image recognition model by adopting an enhanced image in which a latent vector of each element is reflected as learning data. Since the recognition result of each element is verified by the trained AI-based image recognition model, the performance of the AI-based image recognition model trained using an enhanced image in which the latent vector of the element is enhanced is verified. In addition, the training device 220 may iteratively perform the process of training and verifying the AI-based image recognition model until a preset target recognition rate is achieved for each element.

[0055] The determination device 230 compares the recognition rate (or recognition result) of each element included in the enhanced image output from the training device 220 with a preset or defined target recognition rate, and provides a comparison result.

[0056] In this case, the target recognition rate can be set differently according to the autonomous driving level, driving conditions, or elements of the image. The target recognition rate can be set based on the specifications for image recognition of each element, or can be set by an individual or business person who provides the technology.

[0057] In other words, the determination device 230 can judge or determine whether the recognition rate is met based on the specification for each autonomous driving function. When the recognition rate of each element meets the target recognition rate, the process for mass production is performed. When the recognition rate fails to meet the target recognition rate, the relevant enhanced image is transmitted to the analysis device 240.

[0058] When it is determined that an element having a recognition rate lower than the target recognition rate exists based on the comparison result of each element received from the determination device 230, the analysis device 240 may analyze the latent vector of the element having a recognition rate lower than the target recognition rate, and may provide the latent vector of the analyzed element to the conversion device 210 so that the latent vector of the analyzed element is reflected in the generative AI model.

[0059] According to an embodiment, the analyzing device 240 may iteratively provide the latent vectors of the analyzed elements to the converting device 210 until the recognition rate of each element is equal to or greater than the target recognition rate of each element.

[0060] In other words, the analysis device 240 may iteratively perform the following processing: analyzing the latent vector of each element that is not recognized (unrecognized) or is incorrectly recognized (incorrectly recognized) (because the recognition rate of the element fails to meet the target recognition rate in the enhanced image), and providing the analysis results to the conversion device 210 until the recognition rates of all elements meet the target recognition rate.

[0061] As described above, the analysis device 240 may extract elements that provide reasons for under-recognition or misrecognition in an image that fails to meet target performance. The extracted elements may be transferred to a process of extracting a latent vector of an enhanced image or image through an AI model, and the latent vector of each element extracted through the AI ​​model may be analyzed, and a DB may be made using the latent vector.

[0062] In this case, the extracted latent vectors may include latent vectors of factors that affect recognition performance, such as background, shape, texture, tone, and various object types.

[0063] The features extracted by the analysis device 240 may include a latent vector that affects the recognition performance, and the above data (or latent vector) may be used as input data of a conversion device including a generative AI model.

[0064] In this case, the generative AI model can generate advanced enhanced images for improved recognition performance based on information about the input features.

[0065] As described above, AI image recognition models can show improved recognition performance by learning from different and large amounts of data (which are not provided through existing learning / validation cycles) and by augmenting images that include latent vectors that affect recognition performance.

[0066] It is obvious to those skilled in the art that even if the description about the device is omitted, the device according to the embodiment of the present disclosure includes the reference Figures 3 to 6 All data described in Methods.

[0067] In the following, reference will be made to Figures 3 to 6 Operations in an apparatus for learning image recognition according to an embodiment of the present disclosure are described in detail.

[0068] Figure 3 is a flowchart showing the operation of a method for learning image recognition according to another embodiment of the present disclosure, which shows Figure 2 Operational flow of a device for learning image recognition.

[0069] refer to Figure 3 According to another embodiment of the present disclosure, a method for learning image recognition includes: generating an enhanced image by converting an input image (such as a simulated synthetic image) into an enhanced image that enhances specific elements through a generative AI model, for enhancing a preset latent vector of each element (S310).

[0070] In this case, since the input image and feature information are input, the generative AI model can be a model for generating an enhanced image obtained by reflecting the feature information (latent vector) in the input image. The generative AI model can be trained using a previously collected learning dataset.

[0071] In S310, the input image (ie, the synthesized image) is converted into an enhanced image, and the converted enhanced image can be used to train or retrain the AI-based image recognition model. Then, using the AI-based image recognition model, the recognition result of each element included in the enhanced image (eg, the recognition rate of each element) can be provided or output (S320).

[0072] In this case, in S320, the AI-based image recognition model can be trained or retrained by using an enhanced image reflecting the latent vector of each element as learning data, and the recognition result of each element is verified by the trained AI-based image recognition model, thereby verifying the performance of the trained AI-based image recognition model.

[0073] When the recognition result (i.e., recognition rate) of each element is output in S320, the recognition rate of each element is equal to or greater than the target recognition rate (in more detail, autonomous driving level) preset for the element. When an element having a recognition rate lower than the target recognition rate is determined to exist, the latent vector of the element having a recognition rate lower than the target recognition rate is analyzed (S340 and S350).

[0074] Therefore, analyzing the characteristics (latent vector) of each element (S350) may include: Figure 4 As shown, elements with a lower recognition rate are extracted from the enhanced image, and the extracted elements are analyzed based on AI to analyze the characteristics of the extracted elements. Information on the characteristics of the analyzed elements is prepared in the form of a database (DB) reflected in a generative AI model (S410, S420, and S430).

[0075] For example, Figure 5A As shown, the target recognition rate (or based on the recognition rate) can be based on the driving situation (guardrail shadow, building part, or tunnel part) and the element (lane or vehicle; more specifically, the autonomous driving level). Figure 5A In the case of driving in the shadow of the guardrail, the target recognition rates of "lane" and "vehicle" can be 92% and 92% respectively. In addition, in the building part, the target recognition rates of "lane" and "vehicle" can be 91% and 91% respectively. In addition, in the tunnel part, the target recognition rates of "lane" and "vehicle" can be 93% and 90% respectively. Regarding the recognition results of the image recognition model, as shown in Figure 5BAs shown in FIG. 1 , in the driving condition of the guardrail shadow, the recognition results of “lane” and “vehicle” can be 88% and 95% respectively, in the building part, the recognition results of “lane” and “vehicle” can be 93% and 90% respectively, and in the driving condition of the tunnel part, the recognition results of “lane” and “vehicle” can be 94% and 92% respectively. In this case, according to the method of the present disclosure, Figure 6 As shown, the "lane" 620 in the driving condition of the guardrail shadow and the "vehicle" 630 in the driving condition of the building part can be extracted as elements with a recognition rate lower than the target recognition rate. When the features of the extracted elements are analyzed, the latent vector of the element with a lower recognition rate in the enhanced image can be made in the form of DB.

[0076] When the latent vector of an element with a lower recognition rate is analyzed in S350, the input image (i.e., a synthetic image based on simulation) can be converted into another enhanced image in which the analyzed features of the element are enhanced by reflecting the analyzed latent vector of the element in the generative AI model (S360).

[0077] When another enhanced image (such as a new enhanced image) reflecting the analysis features of an element whose recognition rate is lower than the target recognition rate is converted and generated, a process (S320) of training or retraining an AI-based image recognition model using the new enhanced image and verifying the trained AI-based image recognition model and a process (S330 and S340) of comparing the recognition rate of each element in the new enhanced image with the target recognition rate and determining the comparison result.

[0078] The processes of, for example, S320 to S340 and S350 to S360 may be iteratively performed until the recognition rate of each element output by the image recognition model is equal to or greater than the target recognition rate.

[0079] In other words, the method of the present disclosure analyzes the potential vector of each element that is not recognized or misrecognized and cannot meet the target recognition rate in the enhanced image, reflects the information about the analyzed features in the generative AI model, and generates an enhanced image with the analyzed features reflected therein, and it is iteratively performed until the recognition rate of all elements included in the enhanced image meets the target recognition rate. Thus, the AI-based image recognition model can be improved and the AI-based image recognition model can be easily verified.

[0080] It is obvious to those skilled in the art that even if the description about the method is omitted, the method according to another embodiment of the present disclosure includes referring to Figure 2 All information described in the device.

[0081] As described above, according to the apparatus and method for learning image recognition according to the embodiment of the present disclosure, the reason why the recognition rate is not satisfied in the image DB can be analyzed, thereby improving the recognition rate through the image recognition algorithm or the image recognition model.

[0082] In addition, according to the device and method for learning image recognition of the embodiments of the present disclosure, the causes of the decline in recognition performance discovered during the AI-based learning and verification process can be analyzed to extract the features of the image analyzed by AI, thereby regenerating an image reflecting the extracted features, and performing the learning and verification process to improve image recognition performance.

[0083] In addition, according to the device and method for learning image recognition of an embodiment of the present disclosure, when an enhanced image is generated for a synthetic image generated by simulation, the image recognition model is trained by reflecting the latent vector of each element that is not recognized or incorrectly recognized, thereby quantitatively / qualitatively improving the synthetic image, thereby improving the performance of the image recognition model.

[0084] In addition, according to the apparatus and method for learning image recognition of the embodiment of the present disclosure, the reason why the image database (DB) fails to meet the recognition rate can be analyzed. Therefore, the recognition rate of the image recognition algorithm (or model) can be improved. In addition, since a database with substantially real images is constructed, the replacement rate of real images can be increased, so the cost and time for generating images can be reduced.

[0085] In addition, according to the apparatus and method for learning image recognition of the embodiment of the present disclosure, a new DB can be generated by reflecting the potential vector of each element in the existing image, so that the data can be reused. Thus, the diversity of images is enhanced by the scheme of generating images based on AI, thereby firmly performing the training and verification of the image recognition model.

[0086] In addition, according to the apparatus and method for learning image recognition of an embodiment of the present disclosure, the AI ​​model feeds back the features of unrecognized / incorrectly recognized elements through a process having a loop structure set in a closed loop, so that the AI ​​itself enhances the image, and iteratively loops the entire process until the performance meets the target performance, thereby minimizing the involvement of developers in the development process.

[0087] In addition, according to the device and method for learning image recognition of the embodiment of the present disclosure, the AI ​​model overcomes the limitation of the type / range of images generated by humans. Therefore, compared with the conventional development scheme, the coverage of the range required for training and verifying each autonomous driving level can be further expanded and the algorithm can be developed earlier.

[0088] Figure 7 is a block diagram illustrating a method for learning image recognition according to an embodiment of the present disclosure.

[0089] refer to Figure 7 According to another embodiment of the present disclosure, the method for learning image recognition can be implemented by a computing system. The computing system 1000 may include at least one processor 1100, a memory 1300, a user interface input device 1400, a user interface output device 1500, a storage device 1600, and a network interface 1700 connected to each other via a system bus 1200.

[0090] The processor 1100 may be a central processing unit (CPU) or a semiconductor device for processing instructions stored in the memory 1300 and / or the storage device 1600. Each of the memory 1300 and the storage device 1600 may include various types of volatile or non-volatile storage media. For example, the memory 1300 may include a read-only memory (ROM; see 1310) and a random access memory (RAM; see 1320).

[0091] Therefore, the operation of the method or algorithm described in conjunction with the embodiments disclosed in the present disclosure can be directly implemented with a hardware module, a software module, or a combination thereof, and executed by the processor 1100. The software module may reside on a storage medium (i.e., the memory 1300 and / or the storage device 1600), such as a RAM, a flash memory, a ROM, an erasable and programmable ROM (EPROM), an electrical EPROM (EEPROM), a register, a hard disk, a removable disk, or a compact disk ROM (CD-ROM). An exemplary storage medium may be coupled to the processor 1100. The processor 1100 may read information from the storage medium and may write information to the storage medium. Alternatively, the storage medium may be integrated with the processor 1100. The processor 1100 and the storage medium may reside in an application specific integrated circuit (ASIC). The ASIC may reside in a user terminal. Alternatively, the processor 1100 and the storage medium may reside as separate components of the user's terminal.

[0092] According to the present disclosure, the recognition rate of elements to be detected can be improved in an autonomous driving vehicle through an image recognition model based on artificial intelligence (AI).

[0093] According to the present disclosure, the causes of deterioration in recognition performance discovered during an AI-based learning and verification process can be analyzed to extract features of an image analyzed by AI, thereby regenerating an image reflecting the extracted features, and performing a learning and verification process to improve image recognition performance.

[0094] According to the present disclosure, the image recognition model can be trained by reflecting the latent vector of each element that is not recognized or incorrectly recognized when generating an enhanced image for a synthetic image generated by simulation, thereby quantitatively / qualitatively improving the synthetic image, thereby improving the performance of the image recognition model.

[0095] According to the present disclosure, the reason why the image database (DB) fails to meet the recognition rate can be analyzed. Therefore, the recognition rate of the image recognition algorithm (or model) can be improved. In addition, since a database with substantially real images is generated, the replacement rate of real images can be increased, and thus the cost and time for generating images can be reduced.

[0096] According to the present disclosure, a new DB can be generated by reflecting the potential vector of each element in the existing image, so that the data can be reused. Thus, the diversity of images is enhanced by the AI-based image generation scheme, thereby firmly performing the training and verification of the image recognition model.

[0097] Effects produced by the present disclosure are not limited to the above-mentioned effects, and any other effects not mentioned herein will be clearly understood by those skilled in the art to which the present disclosure belongs from the accompanying description.

[0098] In the above, although the present disclosure has been described with reference to exemplary embodiments and the accompanying drawings, the present disclosure is not limited thereto, but various modifications and changes can be made by a technician in the field to which the present disclosure belongs without departing from the spirit and scope of the present disclosure claimed in the attached claims. Therefore, exemplary embodiments of the present disclosure are provided to explain the spirit and scope of the present disclosure, but they are not limited, so that the spirit and scope of the present disclosure are not limited by the embodiments. The scope of the present disclosure should be interpreted based on the attached claims, and all technical concepts within the scope equivalent to the claims should be included in the scope of the present disclosure.

[0099] Although the present disclosure has been described above with reference to exemplary embodiments and the accompanying drawings, the present disclosure is not limited thereto but various modifications and changes may be made by those skilled in the art without departing from the spirit and scope of the present disclosure as claimed in the appended claims.

Claims

1. A device for learning image recognition, the device comprising: a conversion device configured to generate an enhanced image that enhances a specific element by converting the input image into an enhanced image via a generative artificial intelligence (AI) model to enhance a latent vector of at least one element; A training device configured to train an image recognition model based on AI by using the enhanced image; a determining device configured to compare the recognition rate of the at least one element in the enhanced image provided by the image recognition model with a target recognition rate preset for the at least one element; as well as An analysis device is configured to analyze the latent vector of the at least one element having a recognition rate lower than the target recognition rate when the at least one element having a recognition rate lower than the target recognition rate exists in the enhanced image, and provide the analyzed latent vector of the at least one element to the conversion device, so that the latent vector of the at least one element is reflected in the generative AI model.

2. The device according to claim 1, wherein: The conversion device also generates a new enhanced image by reflecting the latent vector of the at least one element analyzed by the analysis device in the generative AI model, and converts the input image into a new enhanced image in which the latent vector of the at least one element analyzed is enhanced.

3. The device according to claim 2, wherein: The determining means also compares the recognition rate of the at least one element in the new enhanced image provided by the image recognition model with the target recognition rate, and Wherein, when the at least one element whose recognition rate is lower than the target recognition rate exists in the new enhanced image, the analysis device also analyzes the latent vector of the at least one element whose recognition rate is lower than the target recognition rate, and provides the latent vector of the at least one element to be analyzed to the conversion device.

4. The device according to claim 3, wherein: The analyzing device also iteratively provides the latent vector of the at least one element being analyzed to the converting device until the recognition rate of the at least one element is equal to or greater than the target recognition rate of the at least one element.

5. The device according to claim 1, wherein: The target recognition rate is set differently according to the autonomous driving level and for the at least one element, and Wherein, the input image includes a simulated synthetic image.

6. A method for learning image recognition, the method comprising: The conversion device generates an enhanced image that enhances a specific element by converting the input image into an enhanced image through a generative artificial intelligence (AI) model to enhance a latent vector of at least one element; A training device trains an image recognition model based on AI by using the enhanced image; The determination device compares the recognition rate of the at least one element provided by the image recognition model with a target recognition rate preset for the at least one element; as well as In response to the at least one element having the recognition rate lower than the target recognition rate existing in the enhanced image, analyzing a latent vector of the at least one element by an analyzing device, and providing the analyzed latent vector of the at least one element to the converting device; as well as The latent vector of the at least one element is reflected in a generative AI model by the conversion device.

7. The method according to claim 6, wherein: Generating the enhanced image comprises: The conversion device generates a new enhanced image by reflecting the latent vector of the at least one element being analyzed in the generative AI model, and the conversion device converts the input image into the new enhanced image in which the latent vector of the at least one element being analyzed is enhanced.

8. The method according to claim 7, wherein: The comparison includes: The determination device compares the recognition rate of the at least one element in the new enhanced image provided by the image recognition model with the target recognition rate, and wherein the analysis comprises: In response to the at least one element having the recognition rate lower than the target recognition rate existing in the new enhanced image, the latent vector of the at least one element is analyzed by the analyzing device.

9. The method according to claim 8, wherein: The generating and reflecting are iteratively performed until the recognition rate of the at least one element is equal to or greater than a target recognition rate of the at least one element.

10. The method according to claim 6, wherein: The target recognition rate is set differently according to the autonomous driving level and for the at least one element, and Wherein, the input image includes a simulated synthetic image.

Images