Method and device for detecting and correcting hallucination of object graph

Abstract

The present disclosure relates to a method and apparatus for detecting hallucinations in an object graph using a hallucination detection unit. According to one embodiment of the present disclosure, a method for detecting hallucinations in an object graph using a hallucination detection unit comprises the steps of: inputting an image and a triplet into a hallucination detection unit; encoding the inputted image and triplet to extract features; inputting the extracted features into a cross-attention block to output a correlation factor; and determining hallucination information based on the output correlation factor.

Description

Technology Field

[0001] The present invention relates to a method and apparatus for detecting and correcting hallucinations in an object graph, and more specifically, to a method and apparatus for detecting and correcting hallucinations that occur during the process of detecting objects from image input and generating a graph using artificial intelligence. By detecting and correcting hallucinations in an object graph based on triplets that constitute objects and relationships, the present invention can be utilized in various industrial applications to obtain more reliable graph data in fields such as image understanding, autonomous driving, visual question answering, and image captioning. Background Technology

[0002] With the recent advancements in computer vision and artificial intelligence technologies, object graph or scene graph generation techniques have been developing to represent visual data, such as images and videos, in a structured form. Object graphs represent objects within an image and their relationships in units of triplets, enabling a higher-dimensional understanding of visual data. These technologies are being utilized in various application fields, including environmental perception in autonomous vehicles, visual question answering (VQA), and image search.

[0003] However, hallucination problems can occur during the object graph generation process due to the incompleteness of input data or limitations of the learning model. Hallucination refers to cases where the object graph contains objects or relationships that do not exist in the actual image. This hallucination problem not only significantly degrades the reliability of the object graph but also negatively impacts the accuracy of application systems based on it.

[0004] Existing object graph generation technologies have lacked mechanisms to effectively address hallucination issues or have limitations in retrospectively modifying generated graphs. Therefore, there is a growing need for systematic methods and devices capable of detecting and correcting hallucination data within object graphs.

[0005] Furthermore, existing technologies aim to generate comprehensive object graphs representing images, which presents a problem in that they are not suitable for applications requiring specific information within images, such as VQA (Visual Question Answering). Additionally, since they are trained on True-Positive data, they are prone to generating False-Positive data corresponding to hallucinations. Moreover, to address these issues, it may be necessary to define hallucination information and perform a negative learning process for the hallucination graph on the dataset; however, redesigning the training process for this purpose entails additional costs, and there is also the problem of insufficient training data volume.

[0006] The aforementioned background technology is technical information that the inventor possessed for the derivation of the present invention or acquired during the process of deriving the present invention, and it cannot be considered as prior art disclosed to the general public prior to the filing of the present invention. The problem to be solved

[0007] The present invention has been devised in consideration of the above-mentioned purpose, and the objective of the present invention is to provide a method and apparatus for detecting and correcting hallucinations in an object graph.

[0008] In addition, the objective of the present invention is to provide a method and apparatus for generating an object graph that has strong resistance to the generation of hallucination information, which trains an artificial intelligence model using a hallucination detector capable of detecting and correcting hallucinations in an object graph or scene graph, and corrects the object graph.

[0009] The problems to be solved in the embodiments of the present invention are not limited thereto, and may also include objectives or effects that can be identified from the means for solving the problems or embodiments described below. means of solving the problem

[0010] A method for detecting hallucinations in an object graph using a hallucination detection unit to solve the problem presented in the present invention may include: inputting an image and a triplet to the hallucination detection unit; encoding the inputted image and triplet to extract features; inputting the extracted features to a cross-attention block to output a correlation factor; and determining hallucination information based on the outputted correlation factor.

[0011] Additionally, the step of encoding the image and triplet to extract features may include the step of inputting the image into an image encoder to extract an image feature map; and the step of inputting the triplet into a triplet encoder to perform vector encoding.

[0012] And, the image encoder is a model that extracts image features containing object classification and relationship information between objects, and the triplet encoder may be a vector encoder for vectorizing natural language.

[0013] Additionally, the step of inputting the above feature into a cross-attention block to output a correlation factor may include: a step of generating an attention map including at least one of a query, a key, and a value; and a step of calculating a correlation score based on the generated attention map.

[0014] In addition, the step of determining the hallucinogenic information may determine that it is not hallucinogenic information if the value of the output correlation factor exceeds a preset threshold, and determine that it is hallucinogenic information if the value of the output correlation factor does not exceed the preset threshold.

[0015] Additionally, a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may further include the step of manipulating at least one value among the subject, relationship information, and object of the triplet; and the step of calculating the similarity between the triplet in which at least one value among the subject, relationship information, and object has been manipulated and the triplet that has not been manipulated.

[0016] And, the step of calculating the similarity above calculates the similarity using a cosine similarity function, and the similarity can be incorporated into the correlation coefficient factor.

[0017] In addition, a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may further include a step of modifying the object graph based on the output correlation factor and the hallucination information.

[0018] And, the step of modifying the object graph may include: a step of replacing at least one relationship information of the triplet; a step of outputting at least one correlation factor based on the replaced relationship information; and a step of modifying the object graph according to the value of the at least one correlation factor output based on the replaced relationship information.

[0019] Additionally, the step of modifying the object graph according to the value of at least one of the correlation factors outputted may include storing the relationship information having the largest value among the at least one correlation factor when at least one of the values ​​of the at least one correlation factor exceeds a preset threshold, and deleting the triplet when none of the values ​​of the at least one correlation factor exceed the preset threshold.

[0020] An apparatus for detecting hallucinations in an object graph using a hallucination detection unit to solve the problem presented in the present invention comprises at least one memory; and at least one processor; wherein the processor inputs an image and a triplet to the hallucination detection unit, encodes the input image and triplet to extract features, inputs the extracted features to a cross-attention block to output a correlation factor, and determines hallucination information based on the output correlation factor.

[0021] The means for solving the above-mentioned problem describes a method and apparatus for detecting hallucinations in an object graph using a hallucination detection unit according to one embodiment of the present invention, and it should be understood that additions, deletions, and changes are possible within the spirit and scope of the invention other than those described above. Effects of the invention

[0022] The present invention can effectively detect hallucinatory phenomena by providing a method and apparatus capable of determining whether a triplet, which is a component of an object graph, is hallucinatory. In particular, the present invention not only determines hallucinations at the triplet level but can also correct incorrect triplet information.

[0023] In addition, the method and apparatus for detecting and correcting hallucinations in an object graph according to the present invention can apply the hallucination detection and correction process not only to object graphs and scene graphs, but also to data of various input forms such as image-graphs, image-captions, and text-graphs.

[0024] In addition, the method and apparatus for detecting and correcting hallucinations in an object graph according to the present invention can process the training of an artificial intelligence model through negative learning using hallucination data. This enables the generation of an object graph with strong resistance to hallucination information, and consequently, prevents the performance of the artificial intelligence model from degrading due to incorrect data. Brief explanation of the drawing

[0025] FIG. 1a is a conceptual diagram showing a system in which various aspects of a method for detecting and correcting hallucinations in an object graph using a hallucination detection unit related to one embodiment of the present invention can be implemented. FIG. 1b is a block diagram illustrating the configuration of a computing device for detecting and correcting hallucinations in an object graph using a hallucination detection unit related to an embodiment of the present invention. Figure 2a is a diagram illustrating object detection for generating an object graph. FIG. 2b is a diagram illustrating a method for extracting objects and relationships to create an object graph. FIG. 2c is a diagram illustrating a method for generating object pairs and relationship information to create an object graph. FIG. 3 is a diagram illustrating a hallucination detection unit for detecting hallucinations using an image and a triplet as inputs according to an embodiment of the present invention. FIG. 4 is a diagram illustrating a system for training an artificial intelligence model by manipulating triplet values ​​using a hallucination detection unit according to an embodiment of the present invention. FIG. 5 is a diagram illustrating a system for generating triplets using image-caption structure data utilizing a hallucination detection unit according to an embodiment of the present invention. FIG. 6 is a diagram illustrating a system for detecting hallucinations using a hallucination detection unit according to an embodiment of the present invention. FIG. 7 is a diagram illustrating a system for generating a modified graph using a hallucination detection unit according to an embodiment of the present invention. FIG. 8 is a diagram illustrating a system for generating a scene graph using a hallucination detection unit according to an embodiment of the present invention. FIG. 9 is a flowchart illustrating a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention. Specific details for implementing the invention

[0026] The following detailed description of the invention refers to the accompanying drawings, which illustrate specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that various embodiments of the invention are different but need not be mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the invention in relation to one embodiment.

[0027] Terms containing ordinal numbers, such as first, second, etc., may be used to describe various components, but the components are not limited by these terms. Terms may be used for the purpose of distinguishing one component from another. For example, it should be understood that a first component may be named a second component, and conversely, a second component may be named a first component.

[0028] The terms used in this application are used merely to describe specific embodiments and are not intended to limit the invention. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this application, terms such as "comprising" or "having" are intended to specify the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0029] Furthermore, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention. Accordingly, the following detailed description is not intended to be taken in a limiting sense, and the scope of the invention is limited only by the appended claims, including all equivalents thereof, provided appropriately described. Similar reference numerals in the drawings refer to the same or similar functions across various aspects.

[0030] FIG. 1a is a conceptual diagram showing a system in which various aspects of a method for detecting and correcting hallucinations in an object graph using a hallucination detection unit related to one embodiment of the present invention can be implemented.

[0031] Referring to FIG. 1a, a system in which various aspects of a method for detecting and correcting hallucinations in an object graph using a hallucination detection unit related to one embodiment of the present invention can be implemented may include at least one of a computing device (10), an external server (20), and a user terminal (30), and each device may mutually transmit and receive data for the system through a network.

[0032] According to one embodiment of the present invention, a computing device (10) or an external server (20) may be a server that provides cloud computing services. More specifically, the computing device (10) or the external server (20) may be a server that provides cloud computing services, which are a type of internet-based computing, where information is processed by another computer connected to the internet rather than the user's computer. A cloud computing service may be a service that stores data on the internet and allows users to access necessary data or programs anytime and anywhere via internet access without installing them on their own computers, and allows data stored on the internet to be easily shared and transmitted through simple operations and clicks. Furthermore, a cloud computing service may not only simply store data on a server on the internet but also allow users to perform desired tasks using the functions of applications provided on the web without installing separate programs, and may be a service that allows multiple people to work while simultaneously sharing documents. Additionally, a cloud computing service may be implemented in at least one form among IaaS (Infrastructure as a Service), PaaS (Platform as a Service), SaaS (Software as a Service), a virtual machine-based cloud server, and a container-based cloud server. That is, the computing device (10) or external server (20) of the present invention may be implemented in at least one form of the cloud computing service described above. The specific description of the cloud computing service described above is merely an example and may include any platform for establishing the cloud computing environment of the present invention.

[0033] According to one embodiment of the present invention, a user terminal (30) may refer to a terminal possessed by a user that receives or can provide information or data related to generating an object graph, such as image information or a triplet, through information exchange with a computing device (10) or an external server (20). For example, the user terminal (30) may be an electronic device for transmitting information related to an input molecule to a computing device (10) or an electronic device for receiving information related to an input molecule processed by a computing device (10).

[0034] According to one embodiment of the present invention, the external server (20) may be a server that stores data for learning deep learning or artificial intelligence. Alternatively, the external server (20) may be a digital device equipped with a processor and memory to have computational capabilities.

[0035] In addition, an external server (20) according to one embodiment of the present invention may have an artificial intelligence, deep learning model, or large language model recorded therein for detecting and correcting hallucinations in an object graph by utilizing a hallucination detection unit. For example, if information or data related to generating an object graph obtained from a user terminal (30) is transmitted to the external server (20), the external server (20) can detect hallucinations in the object graph through an implemented artificial intelligence model.

[0036] Additionally, according to one embodiment of the present invention, the computing device (10), external device (20), and user terminal (30) that build the system may be named interchangeably, and distributed computation may be performed to detect or modify an object graph based on information related to generating an input object graph, or one or more operations may be performed separately.

[0037] FIG. 1b is a block diagram illustrating the configuration of a computing device for detecting and correcting hallucinations in an object graph using a hallucination detection unit related to an embodiment of the present invention.

[0038] According to one embodiment of the present invention, as illustrated in FIG. 1b, a computing device (100) for detecting and modifying hallucinations in an object graph using a hallucination detection unit may include a network unit (110), a memory (120), a sensor unit (130), and a processor (140). The components of the computing device (100) described above are exemplary, and the scope of the present invention is not limited to the components described above. That is, it should be understood that additional components may be included or some of the components described above may be omitted depending on the implementation mode of the embodiments of the present invention.

[0039] According to one embodiment of the present invention, a computing device (100) for detecting and correcting hallucinations in an object graph using a hallucination detection unit may include a network unit (110) capable of transmitting and receiving data with the aforementioned external server (20) and user terminal (30). For example, the network unit (110) may transmit data to the external server (20) for computational processing required by the computing device (100), and may receive data from the external server (20) or the user terminal (30).

[0040] According to one embodiment of the present invention, the memory (120) may store a computer program for performing a method to detect and correct hallucinations in an object graph using a hallucination detection unit, and the stored computer program may be read and executed by a processor (130). Additionally, the memory (120) may store information and / or commands of any form generated or determined by the processor (130) and information of any form received by the network unit (110).

[0041] According to one embodiment of the present invention, the processor (130) can read a computer program stored in memory (120) to perform data processing for learning a deep learning or artificial intelligence model and to perform data processing to detect and correct hallucinations in an object graph by utilizing a hallucination detection unit.

[0042] Figure 2a is a diagram illustrating object detection for generating an object graph.

[0043] In order to generate an object graph, it is necessary to detect objects from an image through object detection processing and obtain information corresponding to each object. Specifically, referring to FIG. 2a, information can be obtained regarding which objects are included in the provided image and what relationships each object has. More specifically, for example, as illustrated in FIG. 2a, the image may include information that a woman is walking down the street wearing pants, information that a woman is walking down the street holding an umbrella, and information that there are windows in the building.

[0044] In addition, object detection can be processed by an artificial intelligence model to generate an object graph. The artificial intelligence model processing object detection may include at least one of a CNN-based artificial intelligence model, a YOLO-based artificial intelligence model, and a DETR-based artificial intelligence model, but is not limited thereto.

[0045] As illustrated in FIG. 2a, information for generating an object graph can be generated by detecting each object included in the image and obtaining information about the object.

[0046] FIG. 2b is a diagram illustrating a method for extracting objects and relationships to create an object graph.

[0047] Referring to FIG. 2b, objects included in the image illustrated in FIG. 2a can be detected and relationships between the detected objects can be modeled to extract object relationships. This method may be referred to herein as a two-step object relationship extraction method.

[0048] Specifically, the two-stage object relationship extraction method detects objects within an image in the first stage and represents the detected objects and their attribute information as nodes in a graph, and based on this, relationship detection can be performed in the second stage. In the second stage, relationships between objects can be defined by utilizing graph structure generation AI that learns the interactions and relationships between each node.

[0049] For example, as illustrated in FIG. 2b, the image in FIG. 2a may have objects "window", "building", "road", "woman", "pants", and "umbrella", and attribute information of the aforementioned objects can be represented as nodes of a graph through object detection. Subsequently, relationships between objects can be defined by utilizing a graph structure generation AI that learns the interactions and relationships between each node. Attribute information of each object, such as "window on building", "woman walking on road", "woman holding umbrella", and "woman wearing pants", and relationships between attribute information can be learned.

[0050] The generated object graph may include subject-relationship-object information as in the example described above, and may be referred to as a triplet.

[0051] Next, referring to FIG. 2c, a triplet consisting of objects and relationship information included in the image illustrated in FIG. 2a can be generated directly. This method may be referred to in this specification as a one-step object relationship extraction method.

[0052] Specifically, the one-step object relationship extraction method can simultaneously extract information on the relationships between objects included in an image. This can be processed through a system that includes an artificial intelligence model configured to simultaneously detect objects and extract relationships between the detected objects.

[0053] For example, objects "window", "building", "road", "woman", "pants", and "umbrella" included in the image of FIG. 2a can be detected, and triplets "window on building", "woman walking on road", "woman holding umbrella", and "woman wearing pants" can be extracted. At the same time, the extracted object information and triplets can be connected to an object graph to generate an object graph.

[0054] FIG. 3 is a diagram illustrating a hallucination detection unit for detecting hallucinations using an image and a triplet as inputs according to an embodiment of the present invention.

[0055] Referring to FIG. 3, a hallucination detection system (300) for detecting hallucinations using an image and a triplet as inputs according to one embodiment of the present invention may include a hallucination detection unit (330) capable of receiving an image input (310) and a triplet input (320) and outputting a correlation factor (350).

[0056] A hallucination detection unit (330) according to one embodiment of the present invention may include an image encoder (332) that encodes an image input (310), an image feature (334) extracted by the image encoder (332), a triplet encoder (336) that encodes a triplet input (320), a triplet feature (338) extracted by the triplet encoder (336), and a cross-attention block (340) that outputs a correlation factor (350) using the extracted image feature (334) and triplet feature (338) as inputs.

[0057] Specifically, the image encoder (332) may be an image feature extraction model for extracting dimensions and features of an image input (310) to extract image features (334). The image feature extraction model may generate an image feature map that the image possesses. The image feature map may include the location, size, and visual features of each object. Additionally, the image feature extraction model may include at least one of a CNN-based artificial intelligence model, a ResNEt-based artificial intelligence model, and a ViT (Vision Transformer)-based artificial intelligence model. However, the image feature extraction model is merely a model for generating an image feature map and is not limited to the examples described above.

[0058] Additionally, the triplet encoder (336) may be a vector encoder for embedding the triplet input (320) into a vector to extract triplet features (338). For example, the vector encoder may include an embedding layer and a layer having a Transformer-based structure to learn text-based relational information.

[0059] Additionally, the cross-attention block (340) can learn the correlation between two types of features by taking the extracted image feature (334) and the extracted triplet feature (338) as inputs. Specifically, it can generate query, key, and value values ​​for each extracted feature and calculate a correlation score based on the generated values. An attention map can be generated while processing the score calculation, and the attention map can be multiplied with the value of the image feature or triplet feature to generate an output that reflects the correlation. The correlation factor (350) may be an output reflecting the correlation generated by performing the cross-attention operation described above. By numerically calculating the relationship between the image and the relationship (text) through the performance of the cross-attention operation described above, the model can understand the relationship between the image and the relationship more precisely.

[0060] FIG. 4 is a diagram illustrating a system for training an artificial intelligence model by manipulating triplet values ​​using a hallucination detection unit according to an embodiment of the present invention.

[0061] Referring to FIG. 4, a system (400) for training an artificial intelligence model by manipulating triplet values ​​using the hallucination detection unit (430) described in FIG. 3 is described.

[0062] Specifically, the graph triplet input (420) can be manipulated into a manipulated triplet (422). For example, the graph triplet input (420) is information consisting of a subject, relationship, and object, and can be manipulated by changing at least one value among the subject, relationship, and object.

[0063] Additionally, the manipulated triplet (422) can be input into the hallucination detection unit (430). As described above, the manipulated triplet (422) can be input into the triplet encoder of the hallucination detection unit (430) to be extracted as a triplet feature. The triplet feature extracted from the manipulated triplet (422) can be processed by the cross-attention operation described above in the cross-attention block with the image feature extracted from the image input (410). A correlation factor (450) can be output by the cross-attention processing.

[0064] Additionally, the manipulated triplet (422) is input to the similarity detection unit (424) together with the graph triplet input (420) so that the similarity between the manipulated triplet (422) and the unmanipulated graph triplet input (420) can be detected. For example, the similarity detection unit (424) can detect similarity by at least one of a cosine similarity function and a dot product similarity function. However, the similarity detection unit (424) is a function for detecting similarity between the manipulated triplet (422) and the unmanipulated graph triplet input (420), and is not limited to the examples described above.

[0065] In addition, in one embodiment of the present invention, the similarity between the manipulated triplet (422) output from the similarity detection unit (424) and the graph triplet input (420) can be integrated into the correlation factor (450) output from the hallucination detection unit (430). Specifically, the calculated similarity (e.g., cosine similarity calculation) can be generated into a similarity matrix, and weights can be assigned to the generated similarity matrix to be integrated into the attention map of the cross-attention block.

[0066] A system (400) for training an artificial intelligence model by manipulating triplet values ​​according to one embodiment of the present invention can generate data from a graph triplet input (420) by changing a value forming an arbitrary triplet (e.g., at least one of a subject, a relationship, or an object). Specifically, hallucinogenic data can be generated by changing a word or a relationship into an antonym based on a vector embedded from the graph triplet input (420). Specifically, data augmentation that increases the amount of data can be performed by changing a word or relationship into a synonym based on a vector embedded from the graph triplet input (420).

[0067] In addition, the system (400) for training artificial intelligence by manipulating triplet values ​​according to one embodiment of the present invention can generate hallucination data by changing arbitrary triplet values, so there is no need to modify the entire object graph that has already been generated to add hallucination information, and the determination of whether or not there is a hallucination can also be easily expressed from the similarity detection unit (424).

[0068] FIG. 5 is a diagram illustrating a system for generating triplets using image-caption structure data utilizing a hallucination detection unit according to an embodiment of the present invention.

[0069] Data with an image-caption structure is a data format configured for learning or analyzing the association between an image and a text caption (e.g., descriptive text of an image). Image caption data can be natural language sentences that describe the content of an image. Specifically, it can be a sentence that explains the concept contained within an image.

[0070] Image caption data can be data that converts the visual information of an image into a linguistic representation. Additionally, image caption data can be data used for Natural Language Processing (NLP) training.

[0071] Referring to FIG. 5, a system (500) for generating a triplet using data of an image-caption structure may include an image input (510), an image caption input (520), a hallucination detection unit (530), and a correlation factor (550).

[0072] The image input (510) can be input to the hallucination detection unit (530) and output as a correlation factor.

[0073] The image caption input (520) may be text data that converts the visual information of the image for the image input (510) into a linguistic expression.

[0074] The graph generation unit (522) can generate a graph triplet input (524) using an image caption input (520) as input.

[0075] The graph triplet input (525) can be input to the hallucination detection unit (530) and output as a correlation factor (550) that indicates a correlation with the image input (510).

[0076] A system (500) for generating triplets using data of an image-caption structure according to one embodiment of the present invention can generate a graph triplet input (524) from an input graph caption input (520) by utilizing a graph generation unit (522). By generating a new triplet from caption data for an image, more diverse learning can be performed cost-effectively.

[0077] Additionally, the generated graph triplet input (524) can be manipulated in the same way as the manipulated triplet (422) shown in FIG. 4 to increase the amount of triplet data and data related to the object graph.

[0078] FIG. 6 is a diagram illustrating a system for detecting hallucinations using a hallucination detection unit according to an embodiment of the present invention.

[0079] A system (600) for detecting hallucinations using a hallucination detection unit according to one embodiment of the present invention can detect whether the data is hallucination data based on a threshold value of a correlation factor (650).

[0080] Specifically, a system (600) that detects hallucinations using a hallucination detection unit (630) can calculate a correlation factor (650) for triplets forming an object graph. If the value of the correlation factor (650) output by the input of each triplet input (620) does not exceed or be greater than a preset threshold, the corresponding triplet input (620) can be defined as hallucination information.

[0081] Additionally, a system (600) that detects hallucinations using a hallucination detection unit (630) can calculate a correlation factor (650) for all triplets (620) forming an object graph and remove triplet inputs (620) defined as hallucination information among all triplet inputs (620).

[0082] FIG. 7 is a diagram illustrating a system for generating a modified graph using a hallucination detection unit according to an embodiment of the present invention.

[0083] Referring to FIG. 7, a system (700) that generates a modified graph using a hallucination detection unit (730) can calculate a correlation factor (750) to modify or remove hallucination information from an object graph for an image. Specifically, if the value of the correlation factor (750) output by the input of each graph triplet input (720) is greater than or equal to a preset threshold and does not exceed, it can be defined as hallucination information.

[0084] Additionally, the system (700) that generates a modified graph using a hallucination detection unit (730) can calculate a correlation factor (750) for all triplets (720) forming the object graph and remove triplet inputs (720) defined as hallucination information among all triplet inputs (720).

[0085] Additionally, the system (700) that generates a modified graph using a hallucination detection unit (730) can determine whether it exceeds or does not exceed a preset threshold and perform an algorithm (760) that modifies the relationship information included in the triplet.

[0086] Specifically, an algorithm (760) that determines whether the relationship information included in the triplet is greater than or less than a preset threshold and modifies the relationship information included in the triplet can, if the value of the correlation factor (750) output by the graph triplet input (720) is greater than or less than the preset threshold, replace the relationship information included in the triplet with all values ​​in the database and recalculate the correlation factor (750).

[0087] Next, if at least one of the correlation factors (750) that are recalculated after all relationship information is replaced exceeds a preset threshold, the relationship information included in the Tricklet is changed and stored as the relationship information calculated with at least one correlation factor (750) that exceeds the preset threshold.

[0088] Additionally, if all values ​​among the correlation factors (750) that have been recalculated with all relationship information replaced do not exceed or are above a preset threshold, it is determined that there is no correlation between the subjects and objects included in the triplet, and the triplet can be deleted.

[0089] After an algorithm (760) is executed to determine whether the relationship information included in the triplet is above or below a preset threshold and to modify the relationship information, a modified graph (770) can be generated in which the relationship information of the triplet is changed or the triplet is deleted.

[0090] FIG. 8 is a diagram illustrating a system for generating a scene graph using a hallucination detection unit according to an embodiment of the present invention.

[0091] Referring to FIG. 8, a system (800) that generates a scene graph using a hallucination detection unit (830) according to one embodiment of the present invention can provide a method for generating an object graph having strong resistance to hallucination information.

[0092] Specifically, when an image input (810) is input into a system (800) that generates a scene graph using a hallucination detection unit (830), an object (or object) can be detected (820) using an artificial intelligence model as described above.

[0093] Based on the detected object (820), an entity list (822) including a group of candidates for subjects and objects corresponding to the information of each object can be generated.

[0094] Next, an algorithm (860) can be performed to determine the correlation factor (850) having the highest value exceeding a preset threshold for all object pairs (e.g., subject and object pairs) detected in the images included in the entity list (822).

[0095] A scene graph (870) can be generated based on object information output from an image and a triplet determined by a correlation factor (850) having the highest value exceeding a preset threshold.

[0096] The scene graph (870) generated by the above-described method may be a two-stage object extraction method described with reference to FIG. 2b.

[0097] The scene graph (870) generated by the above-described method can have strong resistance to hallucinogenic information because the scene graph (870) is generated by selecting a triplet input (824) in which the correlation factor (850) exceeds a preset threshold.

[0098] FIG. 9 is a flowchart illustrating a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention.

[0099] A method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may include the steps of inputting an image and a triplet (S110), extracting features by encoding the image and the triplet (S120), inputting the extracted features into a cross-attention block to output a correlation factor (S130), and determining hallucination information based on the output correlation factor (S140).

[0100] The step of inputting an image and a triplet (S110) may be a step of inputting the image and the triplet into a hallucination detection unit. Specifically, the step of inputting an image and a triplet (S120) may include a step of inputting an image into an image encoder of the hallucination detection unit and a step of inputting a triplet into a triplet encoder of the hallucination detection unit.

[0101] Additionally, the step of extracting features by encoding an image and a triplet (S120) may further include the step of inputting an image into an image encoder to extract an image feature map and the step of inputting a triplet into a triplet encoder to vector encode. Specifically, the image encoder may be an image classification model (e.g., an AI model based on CNN, YOLO, or DNN) to receive an image and extract an image feature map. Specifically, the triplet encoder may be a vector encoder for vectorizing natural language. For example, the triplet encoder may be one of a word embedding-based model, a sentence embedding-based model, a transformer-based multimodal model, a GNN-based model, or a neural scoring-based model.

[0102] Additionally, the step (S130) of inputting the extracted features into a cross-attention block to output a correlation factor may further include the step of generating an attention map containing at least one of a query, a key, and a value, and the step of calculating a correlation score based on the generated attention map.

[0103] And, the step (S140) of determining hallucination information based on the output correlation factor determines whether the value of the output correlation factor exceeds a preset threshold, and if it is above or below the preset threshold but does not exceed it, it is determined to be hallucination information, and if it is above or above the preset threshold but does not exceed it, it is determined not to be hallucination information.

[0104] A method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may further include a step of manipulating a triplet. Specifically, at least one of the subject, relationship information, and object of the triplet may be manipulated. A triplet in which at least one of the subject, relationship information, and object has been manipulated may be input to a similarity detection unit together with an unmanipulated triplet, and the similarity between the manipulated triplet and the unmanipulated triplet may be calculated. The similarity may be integrated into a correlation factor output from the hallucination detection unit.

[0105] In addition, a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may further include the step of inputting an image caption and the step of generating a triplet by a graph generation unit. Specifically, the step of inputting an image caption involves inputting image caption information, which includes a description of the input image, into a graph generation unit, and the graph generation unit may generate a triplet from the image caption. The generated triplet is input into the hallucination detection unit along with the image input to determine whether it is hallucination information.

[0106] In addition, a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may further include a step of modifying the object graph. Specifically, the step of modifying the object graph may include a step of determining whether a preset threshold is exceeded and a step of modifying relationship information included in a triplet. For example, based on an input image and a triplet, the hallucination detection unit outputs a correlation factor, and if the value of the output correlation factor does not exceed a preset threshold, it may be determined as hallucination information and the triplet may be deleted from the object graph.

[0107] In addition, a method for detecting hallucinations in an object graph using a hallucination detection unit according to an embodiment of the present invention may further include a step of generating a scene graph. Specifically, the step of generating a scene graph may include a step of detecting an object based on an input image, a step of creating an entity list of the detected object, and a step of determining a correlation factor or triplet having the highest value exceeding a preset threshold. An object may be detected by using an input image as input to an artificial intelligence model, an entity list may be created based on the detected object, and a scene graph (870) may be generated based on object information output from the image and a triplet determined as the correlation factor (850) having the highest value exceeding a preset threshold based on the created entity list.

[0108] An embodiment according to the present invention may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded on a computer-readable medium. In this case, the medium may include a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical recording medium such as a CD-ROM and a DVD, a magneto-optical medium such as a floptical disk, and a hardware device specifically configured to store and execute program instructions, such as a ROM, RAM, or flash memory.

[0109] Meanwhile, the above-mentioned computer program may be one specifically designed and configured for the present invention, or one known and available to those skilled in the art of computer software. Examples of computer programs may include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0110] According to one embodiment, the method according to various embodiments of the present disclosure may be provided by being included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store (e.g., Play Store™) or directly between two user devices. In the case of online distribution, at least a portion of the computer program product may be temporarily stored or temporarily created in a device-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0111] Although the detailed description of the invention above has been given with reference to preferred embodiments of the invention, those skilled in the art should understand that various modifications and changes can be made to the invention without departing from the spirit and scope of the invention as described in the claimed patent claims.

Claims

Claim 1 A method for detecting hallucinations in an object graph using a hallucination detection unit, comprising: inputting an image and a triplet to the hallucination detection unit; encoding the inputted image and triplet to extract features; inputting the extracted features into a cross-attention block to output a correlation factor; determining hallucination information based on the output correlation factor; and modifying the object graph based on the output correlation factor and the hallucination information, wherein the step of modifying the object graph comprises: replacing at least one relationship information of the triplet based on the correlation factor and the hallucination information; re-outputting the correlation factor based on the triplet in which the at least one relationship information has been replaced; and modifying the object graph by changing and storing the relationship information of the triplet or deleting the triplet according to the value of the re-outputted correlation factor. Claim 2 A method according to claim 1, wherein the step of encoding the image and triplet to extract features comprises: the step of inputting the image into an image encoder to extract an image feature map; and the step of inputting the triplet into a triplet encoder to perform vector encoding. Claim 3 In paragraph 2, the method wherein the image encoder is an image classification model that extracts image features containing object classification and relationship information between objects, and the triplet encoder is a vector encoder for vectorizing natural language. Claim 4 The method of claim 1, wherein the step of inputting the feature into a cross-attention block to output a correlation factor comprises: generating an attention map including at least one of a query, a key, and a value; and calculating a correlation score based on the generated attention map. Claim 5 A method according to claim 1, wherein the step of determining hallucinogenic information is to determine that the output correlation factor is not hallucinogenic information if the value of the output correlation factor exceeds a preset threshold, and to determine that the output correlation factor is hallucinogenic information if the value of the output correlation factor does not exceed the preset threshold. Claim 6 The method of claim 1 further comprises: a step of manipulating at least one value among the subject, relationship information, and object of the triplet; and a step of calculating the similarity between the triplet in which at least one value among the subject, relationship information, and object has been manipulated and the triplet in which it has not been manipulated; wherein the step of calculating the similarity calculates the similarity by a cosine similarity function, and the similarity is integrated into the correlation factor. Claim 7 delete Claim 8 delete Claim 9 delete Claim 10 A device for detecting hallucinations in an object graph using a hallucination detection unit, comprising: at least one memory; and at least one processor; wherein the processor inputs an image and a triplet to the hallucination detection unit, encodes the input image and triplet to extract a feature, inputs the extracted feature to a cross-attention block to output a correlation factor, determines hallucination information based on the output correlation factor, modifies the object graph based on the output correlation factor and the hallucination information, replaces at least one relationship information of the triplet based on the correlation factor and the hallucination information, re-outputs the correlation factor based on the triplet in which the at least one relationship information has been replaced, and modifies the object graph by changing and storing the relationship information of the triplet according to the value of the re-output correlation factor or by deleting the triplet.

Images