Generative ai-based deep learning model optimization apparatus and method for reflecting environmental characteristics using the same

Abstract

The present specification relates to a generative AI-based deep learning model optimization device and a generative AI-based deep learning learning optimization method for reflecting environmental characteristics by the same. The generative AI-based deep learning learning optimization method for reflecting environmental characteristics by the generative AI-based deep learning model optimization device comprises the steps of: collecting images of the surrounding installation environment; generating a condition map based on the characteristics of the installation environment; training a generative AI model with the collected images based on the condition map; generating images to be included in a training dataset through the generative AI model; and optimizing the deep learning model with the generated training dataset.

Description

Technology Field

[0001] This specification relates to Generative AI (hereinafter referred to as Generative AI) technology and the optimization of training of deep learning models. It relates to a technology for generating training data for training a deep learning model using a Generative AI model and effectively training a deep learning model through this. Background Technology

[0002] Generative AI is being used to generate training data for deep learning models and artificial neural networks in various fields. Generative AI generates data based on the user's intent, and the generated data is used to train deep learning models, thereby optimizing the deep learning models.

[0003] Recently, a technology that generates data according to user intent by adding conditions (such as condition maps) to generative AI models (Adding Conditional Control to Text-to-Image Diffusion Models, ICCV 2023) has been introduced. Condition maps used in addition to the user's text input include skeletons and Canny edges. The added conditions have the advantage of providing information that was difficult to express with conventional text input.

[0004] However, since existing deep learning models are generally trained by relying on structured datasets, there is a problem in that they are difficult to reflect the various variability that occurs in the actual installation environment of the device using the deep learning model. In addition, there is a limitation in that optimal performance cannot be achieved in installation environments such as surveillance cameras for CCTV, due to a lack of datasets that match the characteristics or conditions required by the user.

[0005] The above description is intended solely to aid in understanding the background technology regarding the technical concepts of the present invention, and therefore, it should not be understood as prior art known to those skilled in the art of the present invention. The problem to be solved

[0006] This specification is intended to solve the aforementioned problems, and one embodiment of this specification aims to automatically generate an optimized dataset required for training a deep learning model by reflecting the characteristics of the installation environment of a device using a deep learning model, thereby maximizing the performance of the model.

[0007] In addition, one embodiment of the present specification aims to generate training data tailored to the installation environment using generative AI, and to optimize the training process of a deep learning model based on the generated training data to provide highly reliable performance in a real environment.

[0008] The problems that the present invention aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by a person skilled in the art from the description below. means of solving the problem

[0009] The present specification presents a method for optimizing a generative AI-based deep learning model for reflecting environmental characteristics by a generative AI-based deep learning model optimization device to achieve the aforementioned objectives. The method may include the steps of: collecting images of the surrounding installation environment; generating a condition map based on the characteristics of the installation environment; training a generative AI model with the collected images based on the condition map; generating images to be included in a training dataset through the generative AI model; and optimizing the deep learning model with the generated training dataset.

[0010] In addition, the present specification presents a generative AI-based deep learning training optimization device for reflecting environmental characteristics to achieve the aforementioned purpose. The device includes a memory for storing one or more instructions; and a processor for executing the instructions. When the instructions are executed, the processor can collect images of the surrounding installation environment, generate a condition map based on the characteristics of the installation environment, train a generative AI model with the collected images based on the condition map, generate images to be included in a training dataset through the generative AI model, and optimize the deep learning model with the generated training dataset. Effects of the invention

[0011] The embodiments disclosed in this specification have the effect of generating training data tailored to the installation environment using generative AI, and optimizing the training process of a deep learning model based on the generated training data to provide highly reliable performance in a real environment.

[0012] In addition, the embodiments disclosed in this specification have the effect of collecting customized datasets optimized for specific situations.

[0013] In addition, the embodiments disclosed in this specification have the effect of generating training data that systematically reflects various environmental factors by utilizing condition maps in a generative AI model.

[0014] In addition, the embodiments disclosed in this specification have the effect of creating a deep learning model robust to environmental variability, thereby providing high reliability.

[0015] Meanwhile, the effects obtainable from the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art to which the present invention belongs from the description below. Brief explanation of the drawing

[0016] The following drawings attached to this specification illustrate preferred embodiments of the present invention and serve to further enhance understanding of the technical concept of the present invention along with specific details for implementing the invention; therefore, the present invention should not be interpreted as being limited only to the matters described in such drawings. FIG. 1 is a diagram illustrating a method for optimizing a generative AI-based deep learning model for reflecting environmental characteristics according to one embodiment. FIG. 2 is a diagram illustrating the process of collecting images of the surrounding installation environment in a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment. FIG. 3 is a diagram illustrating a method for analyzing environmental similarity in a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment. FIG. 4 illustrates an example of a condition map generation result in a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment. FIG. 5 illustrates the learning structure of a generative AI model in a generative AI-based deep learning model optimization device for reflecting environmental characteristics according to one embodiment. Figure 6 illustrates the process of a condition map generator generating a condition map. Figure 7 illustrates the operation algorithm of the condition map generator. FIG. 8 is a diagram illustrating the process of generating a generative AI training dataset based on installation environment characteristics in a generative AI-based deep learning model optimization device for reflecting environment characteristics according to one embodiment. FIG. 9 is a diagram illustrating the process of generating generative AI data based on variability analysis in a generative AI-based deep learning model optimization device for reflecting environmental characteristics according to one embodiment. FIG. 10 is a block diagram showing a generative AI-based deep learning model optimization device according to one embodiment of the present specification. Specific details for implementing the invention

[0017] It should be noted that technical terms used in this specification are used merely to describe specific embodiments and are not intended to limit the scope of the technology disclosed herein. Furthermore, unless specifically defined otherwise in this specification, technical terms used in this specification shall be interpreted in the sense generally understood by those skilled in the art to which the technology disclosed herein belongs, and shall not be interpreted in an overly broad or overly narrow sense. Additionally, if a technical term used in this specification is an incorrect technical term that fails to accurately express the concept of the technology disclosed herein, it shall be understood as being replaced by a technical term that can be correctly understood by those skilled in the art to which the technology disclosed herein belongs. Furthermore, general terms used in this specification shall be interpreted according to their prior definitions or according to the context, and shall not be interpreted in an overly narrow sense.

[0018] The embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components, regardless of drawing symbols, will be assigned the same reference number, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" for components used in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not have distinct meanings or roles in themselves. Furthermore, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings and should be understood to include all modifications, equivalents, and substitutions that fall within the spirit and technical scope of the invention.

[0019] Terms including ordinal numbers, such as first, second, etc., as used in this specification may be used to describe various components, but said components should not be limited by said terms. Such terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0020] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0021] A singular expression includes a plural expression unless the context clearly indicates otherwise.

[0022] In this application, terms such as “comprising” or “having” are intended to specify the existence 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.

[0023] An image or video according to one embodiment of the present invention includes both still images and video images unless otherwise specifically limited.

[0024] The embodiments disclosed in this specification relate to deep learning training optimization based on generative AI for reflecting environmental characteristics. It proposes a method for optimizing a model by utilizing conditions (such as a condition map) generated by considering the characteristics of the installation environment of a device using a deep learning model, such as an intelligent surveillance camera, for the training of a generative AI model, configuring training data based on images generated through generative AI, and training the deep learning model through this.

[0025] FIG. 1 is a diagram illustrating a method for optimizing a generative AI-based deep learning model for reflecting environmental characteristics according to one embodiment.

[0026] Referring to FIG. 1, a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment can be achieved by performing the process described below by a generative AI-based deep learning model optimization device.

[0027] First, the generative AI-based deep learning model optimization device collects images of the surrounding installation environment (S100). The entire procedure for collecting images of the surrounding installation environment is as illustrated in FIG. 2.

[0028] FIG. 2 is a diagram illustrating the process of collecting images of the surrounding installation environment in a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment.

[0029] Referring to FIG. 2, first, images of the surrounding environment are collected from installed surveillance cameras (S110). At this time, the surveillance cameras capture images in various environments to collect a dataset of various images. Next, similarity analysis of the collected images, that is, analysis of image changes, is performed (S120) to analyze the composition of the similarity space. After performing the analysis of image changes, it is determined whether the diversity of the images is sufficient (S130). If the diversity of the images is sufficient (S130 → Yes), image collection is stopped, and if it is not sufficient (S130 → No), additional images of the surrounding environment are collected through surveillance cameras installed nearby (S140). When collecting images, environmental similarity is analyzed, that is, a similarity space of surrounding images is constructed (S150). There are manual and automatic methods for analyzing environmental similarity. After determining the method for analyzing environmental similarity (S160), if the method is an automatic generation method (S160 → Yes), an image is selected based on a threshold value (S170). If the method is a manual generation method (S160 → No), dimension reduction and visualization are performed (S180). After inputting the user intent (S185), an image is selected (S190).

[0030] The detailed process of the method for analyzing environmental similarity is explained with reference to Fig. 3.

[0031] FIG. 3 is a diagram illustrating a method for analyzing environmental similarity in a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment.

[0032] FIG. 3 describes the visualization results resulting from dimensionality reduction and the automatic and manual methods utilizing them. Referring to FIG. 3, the surveillance area of ​​the surveillance camera of interest (301) is 311, and the surveillance areas of the surrounding surveillance cameras (302, 303, 304) are 312, 313, and 314, respectively. The red dots (350) represent images collected by the surveillance camera of interest (301), and the gray dots (355) represent images collected by the surrounding surveillance cameras (302, 303, 304).

[0033] FIG. 3(a) is a diagram illustrating a manual image selection method, in which the manual method provides the user with the results of similarity spatial analysis through dimensionality reduction and visualization, and selects an image that the user intends to use. At this time, the user sets an input area (320) intended by the user and selects an image that the user intends to use within it. FIG. 3(b) is a diagram illustrating an automatic method, in which the automatic image selection method selects an image that exists within a threshold distance (340) set from the average similarity (330) of an image (red dots (350)) acquired in the installation environment (monitoring area (311)) of the surveillance camera (301).

[0034] Again, referring to FIG. 1, next, the generative AI-based deep learning model optimizer generates a condition map by reflecting the characteristics of the installation environment (S200).

[0035] A generative AI-based deep learning model optimizer analyzes collected images to pre-generate condition maps that reflect key characteristics of the environment and forms a set. These maps can visually represent the variability and features of the installation environment. The elements constituting the condition map include semantic segmentation, Canny edges, depth images, bounding boxes, and skeletons.

[0036] Condition information for generating condition maps is generated based on the optimization model and installation environment. Semantic segmentation, Kenny edges, depth images, object detection, and skeleton information are generated for all image sets as condition maps.

[0037] FIG. 4 illustrates an example of a condition map generation result in a generative AI-based deep learning model optimization method for reflecting environmental characteristics according to one embodiment.

[0038] FIG. 4(a) illustrates an example of a depth image, and FIG. 4(b) illustrates an example of a depth image and object detection. Detected objects are indicated by bounding boxes. FIG. 4(c) illustrates an example of a depth image, object detection, and a skeleton. The skeleton of the detected object indicated by the bounding box is displayed.

[0039] Based on the generated condition information, the condition distribution around the installation environment is defined. The condition distribution is calculated for each element. The semantic segmentation distribution is calculated based on the number of regions and classes, the depth image is calculated based on the number of regions and depth values, and object detection is calculated based on object size, number of objects, and class distribution values. Skeleton information is calculated based on skeleton size, number of objects, and class distribution values.

[0040] Again, referring to FIG. 1, next, the generative AI-based deep learning model optimizer trains a generative AI model (S300).

[0041] Generative AI models have a model structure that generates images using condition maps. Generative AI models utilizing condition maps operate by applying condition maps to general generative AI models to control spatial information in the generated images. Condition maps are generated by a Condition Map Generator. The Condition Map Generator plays the role of dynamically changing the condition maps during the training process.

[0042] FIG. 5 illustrates the learning structure of a generative AI model in a generative AI-based deep learning model optimization device for reflecting environmental characteristics according to one embodiment, and FIG. 6 illustrates the process of a condition map generator generating a condition map.

[0043] The operation process of the condition map generator is identical to the execution procedure of the condition map generator.

[0044] Referring to FIGS. 5 and 6, when a condition map generator generates a condition map (S210), the generated condition map is input into a ControlNet, and the ControlNet trains a generative AI model to generate an output according to the conditions of the condition map (S220). Subsequently, in the condition condition inference step (S230) within the output result, the result of calculating the loss value based on the output of the generative AI model is checked. Based on the loss value calculated for each element, cases where the loss value is higher than an arbitrary threshold are checked. In the condition condition inference result analysis step (S240), for cases where the loss value is higher than the threshold, the condition is checked, and the condition is re-set to reflect the corresponding element in the condition map condition. In the condition map generation condition update step (S250), the condition is reflected so that the condition map is generated with the reset condition.

[0045] Figure 7 illustrates the operation algorithm of the condition map generator.

[0046] The loss function (L) for training a generative AI model is given by Equation 1. In Equation 1, w is the weight, and L is the individual loss function, where w diff L diff ∈ sim L sim is the weight and loss value for the installation environment image generation result, w seg L seg ∈ the weights and loss values ​​for the semantic segmentation condition, w dep Ldep is the weight and loss value for the depth image condition, w obj L obj is the weight and loss value for the object detection condition, w skl L skl represents the weight and loss values ​​for the skeleton.

[0047]

[0048] Mathematical Equation 2 is the installation environment weighted loss function (L sim As representing ), L sim is included in the loss value by applying additional weight to the result of generating the installation environment image. In Equation 2, M i is the mask value (1 for installation environment images, 0 for other images), N data represents an environment image dataset.

[0049]

[0050] Equation 3 is the semantic segmentation condition loss function (L seg As representing ), L seg is defined based on a pixel-based probability distribution according to the number of semantic segmentation regions and classes. In Equation 3, for R regions and C classes, the actual distribution P true (R, C j ) and the generated distribution P pred (R, C j Calculate the difference of ).

[0051]

[0052] Equation 4 is the depth image condition loss function (L dep As representing ), L dep is defined based on the distribution of depth values ​​by region. In Equation 4, for R regions, the depth value D k Regarding the actual depth value distribution P true (R,D k ) and the generated distribution P pred (R,Dk Calculate the difference of ).

[0053]

[0054] Equation 5 is the object detection condition loss function (L obj As representing ), L obj It analyzes object detection conditions and is defined by combining the object size S, the number of objects O, and the distribution of class C; with respect to object size, number, and class, the actual object distribution P true (S,O,C j ) and generated object distribution P pred (S,O,C j Calculate the difference of ).

[0055]

[0056] Equation 6 is the object detection condition loss function (L skl As representing ), L skl It analyzes skeleton conditions, and, identical to object detection conditions, the skeleton object size S s , number of skeleton objects O s , Skeleton Class C s It is defined by combining distributions. With respect to skeleton size, number, and class, the actual object distribution P true (S s ,O s ,C s,j ) and generated object distribution P pred (S s ,O s ,C s,j Calculate the difference of ).

[0057]

[0058] Referring again to FIG. 1, next, the generative AI-based deep learning model optimizer generates a generative AI-based training dataset (S400). The generative AI-based deep learning model optimizer generates a training dataset to be used for optimizing the deep learning model using a generative AI model that has completed training, and this is described in detail below with reference to FIG. 8 and 9.

[0059] FIG. 8 is a diagram illustrating the process of generating a generative AI training dataset based on installation environment characteristics in a generative AI-based deep learning model optimization device for reflecting environment characteristics according to one embodiment.

[0060] Referring to FIG. 8, text information and condition conditions are generated in a generative AI model to generate a dataset required for a target deep learning model (S410), and an image is generated through the generative AI model (S420). Next, the variability of the resulting image generated by the generative AI model is analyzed (S430). It is determined whether the variability analysis condition is satisfied (S440). If the analysis condition is satisfied (S440 → Yes), the corresponding image is saved in the training dataset to update the dataset (S450). If the analysis condition is not satisfied (S440 → No), the process returns to the step of generating the condition conditions (S410).

[0061] FIG. 9 is a diagram illustrating the process of generating generative AI data based on variability analysis in a generative AI-based deep learning model optimization device for reflecting environmental characteristics according to one embodiment.

[0062] Referring to Fig. 9, the detailed generation process of generative AI data is as follows. Condition conditions are generated using installation environment distribution information, and images are generated through a stable diffusion model, which is a generative AI model, based on the generated condition conditions. Next, features of the generated images are extracted using a deep learning-based feature extraction model, and then feature analysis is performed by reducing the dimensionality of the extracted features and analyzing variability relative to the existing installation environment. At this time, a feature boundary, which is a variable region, is set in advance, and an analysis is performed to determine whether the generated image from the generative AI model is included within the pre-set feature boundary. If the generated image is included within the feature boundary, it is included in the dataset.

[0063] Referring again to FIG. 1, finally, the generative AI-based deep learning model optimization device optimizes the deep learning model with the generated dataset (S500). This process is a step of optimizing the deep learning model with the dataset generated by the generative AI model, and at this time, the deep learning model is retrained by setting the learning parameters. During the training process, the performance of the model is periodically evaluated through validation data, and if further performance improvement is required, the process of generating the aforementioned generative AI-based training dataset (S400) is performed again to secure an additional dataset.

[0064] The configuration of a generative AI-based deep learning model optimization device according to one embodiment is described below.

[0065] FIG. 10 is a block diagram showing a generative AI-based deep learning model optimization device according to one embodiment of the present specification.

[0066] Referring to FIG. 10, the generative AI-based deep learning model optimization device (500) is configured to include a communication unit (510), a user interface device (520), a display device (530), a storage medium (540), a processor (550), and a system memory (560), and can perform all functions of the generative AI-based deep learning model optimization device described above.

[0067] The communication unit (510) can transmit and receive signals to and from video recording devices other than the generative AI-based deep learning model optimization device (500) and the interest monitoring camera through the network.

[0068] The user interface device (520) receives user input to control the operations of the generative AI-based deep learning model optimization device (500) or processor (550). The user interface device (520) may include a key pad, a dome switch, a touch pad (pressure / capacitive), a jog wheel, a jog switch, a finger mouse, etc.

[0069] The display device (530) operates in response to the control of the processor (550). The display device (530) displays information processed by the generative AI-based deep learning model optimization device (500) or the processor (550). For example, the display device (530) can display an image according to the control of the processor (550).

[0070] The storage medium (540) may be at least one of flash memory, hard disk, solid state disk type (SSD), multimedia card memory, random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, and optical disk. The storage medium (540) is configured to write and read data in response to the control of the processor (550).

[0071] The processor (550) may include either a general-purpose or dedicated processor and controls the operations of the communication unit (510), user interface device (520), display device (530), storage medium (540), and system memory (560).

[0072] The processor (550) is configured to load program codes containing instructions that provide various functions when executed from the storage medium (540) into the system memory (560) and to execute the loaded program codes. The processor (550) can load an artificial neural network module (561) containing instructions and / or program codes from the storage medium (540) into the system memory (560) and execute the loaded artificial neural network module (561). The artificial neural network module (561) can display images and / or videos received from an external terminal on the display device (530) and detect related user input. Additionally, the artificial neural network module (561) can visualize an additional user interface on the display device (530) and detect user input through it.

[0073] The processor (550) can implement all functions of the generative AI-based deep learning model optimization device (500) described through the artificial neural network module (561) or through a program loaded from the storage medium (540) and all processes of the control method by the generative AI-based deep learning model optimization device (500).

[0074] System memory (560) may be provided as working memory of the processor (550). In the drawing, system memory (560) is shown as a component separate from the processor (550), but this is exemplary and at least a portion of system memory (560) may be integrated within the processor (550). System memory (560) may include at least one of Random Access Memory (RAM), Read Only Memory (ROM), and other types of computer-readable storage media.

[0075] The embodiments described above are combinations of the components and features of the present invention in a specific form. Each component or feature should be considered optional unless otherwise explicitly stated. Each component or feature may be implemented in a form not combined with other components or features. Additionally, it is possible to construct embodiments of the present invention by combining some components and / or features. The order of operations described in the embodiments of the present invention may be changed. Some components or features of one embodiment may be included in another embodiment, or may be replaced with corresponding components or features of another embodiment. It is obvious that embodiments may be constructed by combining claims that do not have an explicit citation relationship in the claims, or that they may be included as new claims through amendments made after filing.

[0076] Embodiments according to the present invention may be implemented by various means, for example, hardware, firmware, software, or a combination thereof. In the case of implementation by hardware, an embodiment of the present invention may be implemented by one or more ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), processors, controllers, microcontrollers, microprocessors, etc.

[0077] In the case of implementation by firmware or software, an embodiment of the present invention may be implemented in the form of a module, procedure, function, etc., that performs the functions or operations described above. The software code may be stored in memory and executed by a processor. The memory may be located inside or outside the processor and may exchange data with the processor by various known means.

[0078] It is obvious to those skilled in the art that the present invention may be embodied in other specific forms without departing from the essential features of the invention. Accordingly, the foregoing detailed description should not be interpreted restrictively in all respects but should be considered exemplary. The scope of the invention shall be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the invention are included within the scope of the invention.

Claims

Claim 1 A method for optimizing a generative AI-based deep learning model for reflecting environmental characteristics using a generative AI-based deep learning model optimization device, comprising: a step of collecting images of the surroundings of an installation environment; a step of generating a condition map based on the characteristics of the installation environment; a step of training a generative AI model with the collected images based on the condition map; and a step of generating images to be included in a training dataset through the generative AI model. The method includes the step of optimizing a deep learning model using the generated training dataset, wherein the elements constituting the condition map include semantic segmentation, depth images, object detection, and skeletons, and the step of generating the condition map includes: generating the condition map based on the characteristics of the installation environment using a condition map generator; inputting the condition map into a ControlNet and training the generative AI model so that the ControlNet generates an output according to the conditions of the condition map; calculating a loss value for each of the elements based on the output of the generative AI model; identifying cases among the loss values ​​that are higher than a predetermined threshold; and resetting the condition map generation conditions to reflect the element corresponding to the higher case among the elements. A method comprising the step of regenerating the condition map by reflecting the re-set condition map generation conditions in the condition map generator, and the step of generating an image to be included in a training dataset through the generative AI model, wherein the step of generating an image by applying the condition map to the generative AI model to generate an image; the step of analyzing whether the generated image is included within a feature boundary, which is a pre-set variable region; and the step of including the generated image in the dataset if the generated image is included within the feature boundary. Claim 2 A method according to claim 1, wherein the step of collecting images of the surrounding installation environment comprises: acquiring images of the installation environment; analyzing changes in the acquired images; determining whether the diversity of the images is sufficient based on the changes in the acquired images; determining whether the diversity of the images is sufficient, and if the diversity of the images is sufficient, no further images are collected, and if the diversity of the images is insufficient, images of the surrounding installation environment are collected; and selecting a training dataset among the collected images according to the result of similarity spatial analysis of the images of the surrounding installation environment. Claim 3 In claim 2, the step of selecting a training dataset among the collected images according to the result of a spatial similarity analysis of images surrounding the installation environment comprises: a step of reducing the dimensionality of the result of the spatial similarity analysis and visualizing it; a step of setting a user intent input area in the installation environment; and a step of selecting an image within the user intent input area as the training dataset. Claim 4 In claim 2, the step of selecting a training dataset among the collected images according to the result of spatial analysis of the similarity of images around the installation environment comprises: a step of calculating the average similarity of images acquired in the installation environment; and a step of selecting an image within a predetermined value from the average similarity as the training dataset. Claim 5 Memory that stores one or more instructions; and a processor for executing the above instructions; wherein, when the above instructions are executed, the processor is configured to collect images of the surrounding installation environment, generate a condition map based on the characteristics of the installation environment, train a generative AI model with the collected images based on the condition map, generate images to be included in a training dataset through the generative AI model, and optimize a deep learning model with the generated training dataset, wherein the elements constituting the condition map include semantic segmentation, depth images, object detection, and skeletons, wherein, in the process of generating the condition map, the processor uses a condition map generator to generate the condition map based on the characteristics of the installation environment, inputs the condition map into a ControlNet, and proceeds with training the generative AI model so that the ControlNet generates an output according to the conditions of the condition map, calculates a loss value for each of the elements based on the output of the generative AI model, and among the loss values, cases higher than a predetermined threshold value A generative AI-based deep learning training optimization device for reflecting environmental characteristics, configured to verify and, in order to reflect the element corresponding to the high case among the above elements, reset the condition map generation conditions and reflect the reset condition map generation conditions in the condition map generator to regenerate the condition map, and the processor, in the process of generating an image to be included in the training dataset through the generative AI model, applies the condition map to the generative AI model to generate an image, analyzes whether the generated image is included within a feature boundary which is a pre-set variable region, and if the generated image is included within the feature boundary, includes the generated image in the dataset. Claim 6 In claim 5, the processor acquires an image of an installation environment, analyzes changes in the acquired image, determines whether the diversity of the image is sufficient based on changes in the acquired image, and if the diversity of the image is sufficient as a result of determining whether the diversity of the image is sufficient, no further images are collected, and if the diversity of the image is not sufficient, images around the installation environment are collected, and a training dataset is selected among the collected images according to the results of spatial analysis of similarity of images around the installation environment, a generative AI-based deep learning training optimization device for reflecting environmental characteristics. Claim 7 In claim 6, the processor reduces the dimensionality of the similarity spatial analysis result and visualizes it, sets a user intent input area in the installation environment, and selects an image within the user intent input area as the training dataset, a generative AI-based deep learning training optimization device for reflecting environmental characteristics. Claim 8 In claim 6, the processor calculates the average similarity of images acquired in an installation environment and selects images within a predetermined value from the average similarity as the training dataset, a generative AI-based deep learning training optimization device for reflecting environmental characteristics.

Images