Abstract: A system may identify a defined rectangular picture area and render video corresponding to the defined rectangular picture area. The system may receive a video bitstream comprising a picture having a header and may receive data specifying a structure of the picture. The system may parse the data specifying the structure of the picture for an identifier corresponding to a defined rectangular area in the first picture and for a tile index of a top left tile in the defined rectangular area. The system may determine one or more tiles comprised in the defined rectangular area based on the identifier corresponding to the defined rectangular area and the tile index of the top left tile. The system may reconstruct the picture including a sub-picture that comprises the defined rectangular area based upon the identifier corresponding to the defined rectangular area. The computing system may render the sub-picture in the defined rectangular area.

Abstract: Disclosed herein are system, method, and computer program product embodiments for surface automation in black box environments. An embodiment operates by determining scenarios of an application for automation; detecting the scenario during an execution of an application; capturing and storing one or more user interface screenshots of the scenario; identifying and storing user interface information from the user interface screenshot; implementing a sequential set of instructions comprising at least one textual element detection technique and at least one non-textual element detection technique; and executing the sequential set of instructions.

Abstract: A method may include providing one or more inputs to a hybrid data generator, wherein one of the one or more inputs is based at least in part on a wellsite location, wherein the hybrid data generator comprises a large language model, and wherein the large language model is based at least in part on a machine learning algorithm. The method may further include utilizing an information handling system to generate a drilling program based at least in part on the one or more inputs and the hybrid data generator. The method may further include performing at least a portion of a drilling operation based at least in part on the drilling program and collecting at least one measurement from at least one sensor during the drilling operation.

Abstract: A method for monitoring protected regions of a camera view includes receiving, at a processor of a video camera system, a video stream including a series of video frames depicting at least one scene. A background model is identified based on a predefined representation of at least one region of interest associated with the scene. Each of the video frames is compared to the background model, to define an associated binary mask. A persistent change in the at least one region of interest is detected, based on the binary masks, the persistent change having an associated duration that exceeds a predefined threshold. In response to detecting the persistent change, a change notification message including a video file is generated. The method also includes causing display of the change notification message via a user interface and/or causing transmission of the change notification message to a remote compute device.

Abstract: A method and an apparatus for data retrieval in a lightfield database are described. An acquiring unit of the apparatus acquires a sample data. An operation unit determines a plurality of two-dimensional (2D) images from each 4D lightfield data of the lightfield database, and match the sample data with the determined 2D images according to geometrical parameters of the sample data and geometrical parameters of the 2D images. One of the 2D images is selected as a target image based on the matching result, and the 4D lightfield from which the target image is determined is retrieved.

Abstract: In various example embodiments, machine vision-based techniques for non- contact SHM are provided that may integrate both phase-based algorithms and template matching algorithms to enable selection of one or more machine vision algorithms that are effective at measuring responses (e.g., displacement, strain, acceleration, velocity, etc.) under present conditions. Results of a single algorithm or a combination of results from multiple algorithms may be returned. In such techniques, improved template matching algorithms may be employed that provide sub-pixel accuracy. Responses may be adjusted to cancel out camera vibration and video noise. Defects or damage may be determined by tracking changes in displacement within an area of interest.

Abstract: A method for simulating a technical system. Time series are obtained with the aid of a simulation model of the system, variable values being assigned to at least one epistemic parameter of the simulation model, at least one measurement series is obtained by corresponding measurements on the system, for each value of the epistemic parameter, a real-value error measure of the time series obtained for this value with respect to the measurement series and a distribution function of the error measure are calculated, and for the simulation, that value is used, for which the distribution function has the smallest distance to a Heaviside function.

Abstract: This invention is a system and method for utilizing artificial intelligence to operate a surgical robot (e.g., to perform a laminectomy), including a surgical robot, an artificial intelligence guidance system, an image recognition system, an image recognition database, and a database of past procedures with sensor data, electronic medical records, and imaging data. The image recognition system may identify the tissue type present in the patient and if it is the desired tissue type, the AI guidance system may remove a layer of that tissue with the end effector on the surgical robot, and have the surgeon define the tissue type if the image recognition system identified the tissue as anything other than the desired tissue type.

Abstract: According to one embodiment, a defect inspection apparatus includes a storage and a processor. The storage stores dictionary in which a first design image generated by software is associated with a first real image corresponding to the first design image. The processor acquires a second design image generated by the software. The processor searches for a similar first design image similar to the second design image. The processor generates a reference image that is a pseudo real image of a second inspection object, based on the second design data and free of defect, by using the first real image associated with the similar first design image. The first design image includes information of an image area wider than the first real image.

Abstract: A method for blur correction determines a first constant for a first pixel segment and a second constant for a second pixel segment for each of the multipliers that results in a first mean error for the first pixel segment being and a second mean error for the second pixel segment being within a mean error threshold. A first multiplier is selected from the multipliers with the corresponding first constant for the first pixel segment and second multiplier is selected from the multipliers with the corresponding second constant for the second segment that has the probability of the frames with the measured ranges being within the selected error threshold. A data file comprising the first multiplier, the corresponding first constant for the first multiplier, the second multiplier, and the corresponding second constant for the second multiplier is created for a platform computer system for a platform.

Abstract: An airflow detection apparatus (1) includes an airflow detection apparatus (1) that detects an airflow in a painting booth, and includes a thread-like member (2) installed inside the painting booth, an image capture device (4) capable of generating image data obtained by capturing an image of the thread-like member (2), and a calculation device capable of performing calculation processing on the image data. The thread-like member (2) is installed in the painting booth in a mode in which at least a portion thereof is suspended. The calculation device specifies a portion corresponding to the thread-like member (2) in the image data, and detects an airflow based on a position of the portion.

Abstract: Systems and methods for presenting information, related to products or services being shown on a second display device, on a first display device are disclosed. Exemplary implementations may: capturing a marker, by a first display device, displayed on a second display device, decoding information presented in the marker, receiving information related to the product or service being displayed on the second display device; and based on the information, displaying a three dimensional (3D) rendition of the product or service on the first display device.

Abstract: A method includes defining a geometric capsule that is interpretable by a capsule neural network, wherein the geometric capsule includes a feature representation and a pose. The method also includes determining multiple viewpoints relative to the geometric capsule and determining a first appearance representation of the geometric capsule for each of the multiple viewpoints. The method also includes determining a transform for each of the multiple viewpoints that moves each of the multiple viewpoints to a respective transformed viewpoint and determining second appearance representations that each correspond to one of the transformed viewpoints. The method also includes combining the second appearance representations to define an agreed appearance representation. The method also includes updating the feature representation for the geometric capsule based on the agreed appearance representation.

Abstract: Provided are a method and an apparatus for generating image data for machine learning. The method may include extracting first image data for a product generated as a result of conducting a first step among a plurality of steps of a manufacturing process of a product based on 3D CAD data for the product, generating a plurality of posture conversion image sets corresponding to a plurality of postures for the product, respectively, based on the first image data, performing image processing through a rendering engine by applying actual image data to respective images of the plurality of posture conversion image sets, and generating a learning data set for the first image data through the image processing.

Abstract: Systems and methods are disclosed for high dynamic range anti-ghosting and fusion. For example, methods may include receiving images from image sensors in a linear domain, each image having different exposures or gains, blending luminance values at each pixel from each of the images to generate a blended image, selecting an useful image based on degree of useful information for a pixel, calculating a distance value from the images for the pixel, locating from a look-up table (LUT) an anti-ghosting weight using the useful image and the distance value for the pixel, proportionally applying the located anti-ghosting weight to the pixel for each of the input images to generate an output image, all being performed in the linear domain, and storing, displaying, or transmitting the output image based on at least the anti-ghosting weight.

Abstract: There is described a deep learning supervised regression based model including methods and systems for facial attribute prediction and use thereof. An example of use is an augmented and/or virtual reality interface to provide a modified image responsive to facial attribute predictions determined from the image. Facial effects matching facial attributes are selected to be applied in the interface.

Abstract: A method for determining similarity of adjacent rows in a picture, the method includes obtaining feature value of corresponding sub-pixel of each row of picture data in a frame of picture data; subtracting feature values of corresponding sub-pixels of two adjacent rows of picture data in the frame of picture data, and summing absolute values of multiple differences obtained from subtracting; and determining similarity of the two adjacent rows of picture data in the frame of picture data according to the summed value. The present application further provides a display device, the display device includes an apparatus for determining two adjacent rows in a picture, the apparatus includes processor and memory for implementing steps of the above methods.

Abstract: A method of classifying a media includes receiving a media file and extracting therefrom first and second data streams including first and second media content, respectively, the media content being associated with the media item. First and second feature vectors describing the first and second media content, respectively, are generated. At least a first single feature vector representing the first sequence of first feature vectors and the second sequence of second feature vectors is generated, or at least a first single feature vector representing at least the first sequence of first feature vectors is generated. A second feature vector representing at least the second sequence of second feature vectors is generated. A probability vector from the first single feature vector, or from the first and second single feature vectors is generated. A user profile suitability class is assigned to the media item based on the probability vector.

Abstract: Provided herein are systems and methods for extracting text from a document. Different optical character recognition (OCR) tools are used to extract different versions of the text in the document. Metrics evaluating the quality of the extracted text are compared to identify and select higher quality extracted text. A selected portion of text is compared to a threshold to ensure minimal quality. The selected portion of text is then saved. Error correction can be applied to the selected portion of text based on errors specific to the OCR tools or the document contents.

Abstract: Disclosed are a method and a device for marking an image position of a sub-pixel of a display screen, and a storage medium. The method includes: obtaining an original image of the sub-pixel of the display screen that has been taken, marking all the sub-pixel bright points in the original image of the sub-pixel to obtain a sub-pixel mark map; searching for points to be filled among adjacent points of the bright point of each sub-pixel in the sub-pixel mark map according to the first direction pixel pitch and the second direction pixel pitch in the original image of the sub-pixel, filling the points to be filled to obtain a filling mark map; and obtaining the target sub-pixel mark map based on the sub-pixel mark map and the filling mark map.

Abstract: A dishwasher (1), in particular a commercial dishwasher (1), which is in the form of a box-type dishwasher and has a treatment chamber (2) which can be closed and in which washware can be treated in accordance with a predefined or predefinable treatment program, wherein a washware detector apparatus (51) is designed to detect the type of washware to be, wherein the washware detector apparatus (51) has an optical identification system, which is oriented toward the washware to be treated, and an evaluation device, wherein the optical identification system is designed to record at least a portion of the light reflected by the washware to be treated as a reflection image, and wherein the evaluation device is designed to automatically distinguish between different types of washware on the basis of the recorded reflection image, wherein the extent of the reflection provides information about the type of washware.

Abstract: Disclosed are an image matching method and apparatus. The image matching method is inclusive of steps of obtaining a panoramic image of at least one subspace in a 3D space and a 2D image of the 3D space; acquiring a 2D image of the at least one subspace in the 3D space; performing 3D reconstruction on the panoramic image of the at least one subspace, and procuring a projection image corresponding to the panoramic image of the at least one subspace; and attaining a matching relationship between the panoramic image of the at least one subspace and the 2D image of the at least one subspace, and establishing an association relationship between the panoramic image of the at least one subspace and the 2D image of the at least one subspace between which the matching relationship has been generated.

Abstract: An apparatus for estimating a camera pose according to an embodiment of the present disclosure includes a similar image searcher, a clusterer, and an estimator. The similar image searcher searches for a plurality of images similar to an input image, from among a plurality of previously-stored images, based on the input image. The clusterer creates a cluster including at least some similar images meeting predetermined conditions, from among the plurality of similar images, based on viewpoint data tagged to each of the plurality of similar images. The estimator estimates a pose of a camera that has generated the input image, based on the cluster.

Abstract: Multimodal techniques are described for classifying start pages and document types of an unstructured document image package. To that end, some implementations of the disclosure relate to a method, including: obtaining a document image file including multiple pages and multiple document types; generating, for each page of the multiple pages, using multiple independent trained models, multiple independent predictions, each of the multiple independent predictions indicating: whether or not the page is a first page of a document, or a document type of the multiple document types that the page corresponds to; and generating, for each page of the multiple pages, based on the multiple independent predictions, using a neural network, a final prediction output indicating whether or not the page is the first page of a document, or one of the multiple document types that the page corresponds to.

Abstract: In various embodiments, a match cutting application automatically generates candidate match cuts. In some embodiments, the match cutting application detects shot sequences within one or more video sequences, deduplicates the shot sequences, computes an encoding of each shot sequence after the deduplication, and compares the encodings associated with pairs of shot sequences using a distance metric to identify pairs of shot sequences that can be cut together to generate candidate match cuts. In addition, the match cutting application can apply an approximation technique, such as an approximate nearest neighbor search technique, to identify intermediate pairs of shot sequences prior to comparing encodings associated with the intermediate pairs of shot sequences.

Abstract: An inspection device includes: an electromagnetic wave irradiator that irradiates, with a predetermined electromagnetic wave from a first film side, the package that is conveyed along a predetermined direction and that has the spaces at a plurality of positions in a width direction; an imaging device that is disposed opposed to the electromagnetic wave irradiator across the package, includes an electromagnetic wave detector including a plurality of detection elements that is arrayed along the width direction and that detects the electromagnetic wave radiated from the electromagnetic wave irradiator and transmitted through the package, and sequentially outputs an obtained electromagnetic wave transmission image every time the package is conveyed by a predetermined amount; and an image processing device that processes an image signal output from the imaging device.

Abstract: A collation section (14) extracts a difference between two pieces of operation setting data which are pieces of data each of which is for setting operation in a device, and generates difference information indicating the extracted difference, before a user requests. A collation result management section (15) stores in a storage device (150), the difference information generated by the collation section (14). A display section (16) acquires the difference information from the storage device (150) and outputs the acquired difference information to a display device when the user requests.

Abstract: A multi-model medical scan analysis system is operable to generate a plurality of training sets from a plurality of medical scans. Each of a set of sub-models is generated by performing a training step on a corresponding one of the plurality of training sets of the plurality of medical scans. A set of abnormality data is generated by applying a subset of a set of inference functions on a new medical scan. The subset of the set of inference functions utilize the subset of the set of sub-models, and each of the set of abnormality data is generated as output of performing one of the subset of the set of inference functions. The multi-model medical scan analysis system is further operable to generate final abnormality data that includes a global probability indicating a probability that any abnormality is present based on the set of abnormality data.

Abstract: A location of an object of interest (205) is determined using both observations and non-observations. Numerous images (341-345) are stored in a database in association with image capture information, including an image capture location (221-225). Image recognition is used to determine which of the images include the object of interest (205) and which of the images do not include the object of interest. For each of multiple candidate locations (455) within an area of the captured images, a likelihood value of the object of interest existing at the candidate location is calculated using the image capture information for images determined to include the object of interest and using the image capture information for images determined not to include the object of interest. The location of the object is determined using the likelihood values for the multiple candidate locations.

Abstract: A digital image processor includes a region proposal network configured to transform digital image inputs into region proposals and bounding box refinement logic configured to transform the region proposals by determining a first set of the region proposals exhibiting dense spacing, determining a second set of the region proposals exhibiting sparse spacing, executing a first transformation to merge at least some of the region proposals exhibiting dense spacing to generate refined region proposals, executing a second transformation to join at least some of the region proposals exhibiting sparse spacing to generate additional ones of the refined region proposals, and applying an expansion transformation to the refined region proposals.

Abstract: A depth sensing camera system that comprises one or more fisheye lenses and infrared and/or near-infrared image sensors. In some examples, the image sensors may generate output signals based at least in part on receiving radiation via the fisheye lenses. A depth measurement may be calculated based at least in part on the output signals. For example, these output signals may be provided as input to a depth model, which may determine the depth measurement. In some examples, such a depth model may be integrated into an application-specific integrated circuit and/or may be operated by an application processor.

Abstract: Methods and systems are disclosed for performing operations comprising: receiving an image that includes a depiction of a person wearing a fashion item; generating a segmentation of the fashion item worn by the person depicted in the image; identifying a facial expression of the user depicted in the image; and in response to identifying the facial expression, applying one or more augmented reality elements to the fashion item worn by the person based on the segmentation of the fashion item worn by the person.

Abstract: A performing device of an impulse-like gesture recognition system executes an impulse-like gesture recognition method. A performing procedure of the impulse-like gesture recognition method includes steps of: receiving a sensing signal from a sensing unit; determining a prediction with at least one impulse-like label according to the sensing frames by a deep learning-based model; and classifying at least one gesture event according to the prediction. The gesture event is classified to determine the motion of the user. Since the at least one impulse-like label is used to label at least one detection score of the deep learning-based model, the detection score is non-decreasing, reaction time of the at least one gesture event for an incoming gesture is fast, rapid consecutive gestures are easily decomposed, and an expensive post-processing is not needed.

Abstract: A method is disclosed for indexing 3D digital models, retrieving them, comparing them and displaying the results in a 3D space. The method comprises four complementary parts, i.e. displaying, comparing/searching, reconciling the faces, and classifying the results. These parts can overlap with each other or can be implemented separately. A method is described for retrieving 3D models that share certain similarities of form with a reference 3D model, involving a first step of analysis in order to generate representations (descriptors). The process of searching/comparing 3D models based on descriptors partially related to the faces optionally requires a process of pairing and reconciling the faces. The results are displayed in a single 3-dimensional space and, owing to a mark on the faces of the 3D models, makes it possible to distinguish several types of difference between similar 3D models.

Abstract: Aspects of the disclosure relate to detecting sidewalks adjacent to roads. In this regard, a set of potential sidewalk areas adjacent to one or more roads in a vehicle's vicinity may be determined based on map data. Topology data for the set of potential sidewalk areas may be generated based on sensor data received from a perception system of the vehicle. The set of potential sidewalks may be filtered to remove areas unlikely to include a sidewalk. The vehicle may be operated based on the filtered set of potential sidewalk areas, which may include taking precautionary measures when within a predetermined distance from any of the filtered set of potential sidewalks.

Abstract: A face search method and apparatus are provided. The method includes obtaining a to-be-searched face image, and inputting the face image into a first feature extraction model to obtain a first face feature. The method further includes inputting the face image and the first face feature into a first feature mapping model for feature mapping, to output a standard feature corresponding to the first face feature, and performing face search for the face image based on the standard feature. Features extracted by using a plurality of feature extraction models are concatenated, and a concatenated feature is used as a basis for constructing a standard feature.

Abstract: A region extraction device acquires a first frame image and a second frame image which are temporally successive. The device acquires, based on the acquired first frame image and second frame image, motion information indicating a region where a motion exists in the first frame image, detects, based on the acquired first frame image, positions of an elbow and a wrist of a human body from the region where the motion exists indicated by the acquired motion information, and extracts, based on the detected positions, a region corresponding to a part, of the region where the motion exists indicated by the acquired motion information, located in a human body hand's side of the wrist.

Abstract: Provided are a photographing method and a terminal, relating to the field of electronic devices, and by which a user can be instructed based on a template to adjust a position or an attitude of a mobile phone, to obtain a photo or a video whose composition is the same as or similar to that of the template, thereby improving the user photographing experience. The method specifically includes: after starting a first function of a camera application, an electronic device calculates a moving direction and distance of the electronic device and/or calculates a rotation direction or a rotation angle of the electronic device based on a target template selected by a user and a currently acquired preview and through feature point matching or an AR repositioning technology, so that the electronic device can photograph a photo or a video whose composition is the same as that of the target template.

Abstract: An image selector selects images for visual similarity in some aspects and visual distinction in other aspects. A user submits a query to a search engine. Images identified by the search engine as relevant to the query are clustered into clusters having at least one similar attribute. Selection mechanisms organize images selected from each cluster based on a similarity metric into image rows. A selection metric is calculated for each image row and a row to be displayed is selected based on the selection metric. In some embodiments, images representative of a subset of image rows are displayed to a user and responsive to selection of a representative image, the corresponding image row is displayed.

Abstract: Authentication of electronic document is based on multiple digital signatures incorporated into a blockchain. Structured data, metadata, and instructions may be hashed to generate the multiple digital signatures for distribution via the blockchain. Any peer receiving the blockchain may then verify an authenticity of an electronic document based on any one or more of the multiple digital signatures incorporated into the blockchain.

Abstract: Systems and methods for displaying augmented reality content relative to an identified object are disclosed. The systems and methods can include a set of operations. An operation can include providing a first image. An operation can include determining interest points in the first image. An operation can include identifying an object in the first image. An operation can include identifying augmented reality content associated with the object. An operation can include determining a first transformation for displaying the augmented reality content in the first image relative to the identified object. An operation can include providing the interest points and the first transformation. An operation can include determining a second transformation for displaying the augmented reality content in a second image relative to the identified object using, at least in part, the first transformation and the interest points. An operation can include displaying, by a user device, the augmented reality content.

Abstract: Methods, apparatuses, and non-transitory machine-readable media for image location based on a perceived interest and display position. Apparatuses can include a display, a memory device, and a controller. an example controller can assign a perceived interest and sort images based in part on the perceived interest. In another example, a method can include assigning, by a controller coupled to a memory device, a perceived interest to an image of a plurality of images, wherein the perceived interest is assigned based in part on a change in position of a display coupled to the memory device while the image is viewable on the display, selecting the image from an initial viewing location on the display responsive to the assigned perceived interest, and transferring the image to a different viewing location, wherein the initial viewing location and the different viewing location are visible on the display.

Abstract: Disclosed are examples of a system, techniques and non-transitory computer-readable storage medium that enable the selection of a response document template for generating an appropriate response to notification document. Respective components may generate an electronic document from the notification document. A common theme between two or more of the electronic documents may be identified. The electronic documents identified as having the common theme may be grouped in a notification document corpus. The notification document corpus may be analyzed with reference to received notification documents stored in a document repository. Based on results of the analysis, a template of a notification response document appropriate for responding to a specific notification document received by the enterprise may be selected. The selected response document template may be forwarded to a client device for further processing.

Abstract: A model including an information extraction layer that obtains image information of a training object in a depth image; a pixel point positioning layer that performs position estimation on a three-dimensional coordinate of human-body key points, defines a body part of the training object as a body component, and calibrates a three-dimensional coordinate of all human-body key points corresponding to the body component; a feature extraction layer that extracts a key-point position feature, a body moving speed feature, and a key-point moving speed feature for action recognition; a vector dimensionality reduction layer that combines the key-point position feature, the body moving speed feature, and the key-point moving speed feature as a multidimensional feature vector, and performs dimensionality reduction on the multidimensional feature vector; and a feature vector classification layer that classifies the multidimensional feature vector that is performed with dimensionality reduction, to recognize a fitness act

Abstract: Examples methods, apparatus/systems and articles of manufacture for auditing point-of-sale images are disclosed herein. Example methods disclosed herein include comparing a region of interest of an image displayed via a user interface with a plurality of reference product images stored in a database to identify a plurality of candidate product images from the plurality of reference product images as potential matches to a first product depicted in the image. For example, the candidate product images are associated with respective confidence levels indicating respective likelihoods of matching the first product. Disclosed example methods also include displaying, via the user interface, the candidate product images simultaneously with the image in a manner based on the respective confidence levels, and automatically selecting a first one of the candidate product images as matching the first product based on the respective confidence levels.

Abstract: A private information detector for data loss prevention is described.

Abstract: A computer-implemented method for computing a prediction on images of a scene includes: receiving one or more polarization raw frames of a scene, the polarization raw frames being captured with a polarizing filter at a different linear polarization angle; extracting one or more first tensors in one or more polarization representation spaces from the polarization raw frames; and computing a prediction regarding one or more optically challenging objects in the scene based on the one or more first tensors in the one or more polarization representation spaces.

Abstract: Aspects of the disclosure are directed to an artificial reality (XR) browser configured to trigger an immersive experience. Implementations display an element at a browser chrome of the XR browser when an immersive experience is loaded. For example, the XR browser can include an application programming interface (API) that supports configuration of a browser chrome element by components of a webpage. The API call can cause the display of the browser chrome element, change a display property for the browser chrome element, or configure the browser chrome element in any other suitable manner. Upon receiving input at the browser chrome element (configured by the API call), the XR browser can transition from displaying a two-dimensional panel view of a webpage supported by loaded web resources (e.g., hypertext transfer protocol (HTTP) pages, graphic images, etc.) to a three-dimensional environment supported by preloaded immersive resources (e.g., three-dimensional models, graphic images, etc.).

Abstract: Disclosed is a method including presenting an input field on a user interface of a generalized search entity, wherein the generalized search entity processes data using a generalized search engine that indexes and searches both merchant sites and non-merchant sites, receiving user input in the input field and determining whether the user input corresponds to a product in a product database to yield a determination. When the determination indicates that the user input does correspond to the product in the product database, the method includes presenting a purchase-related search result, wherein the purchase-related search result is configured such that when a user interacts with the purchase-related search result and confirms a purchase associated with the purchase-related search result, the generalized search entity initiates a purchasing process for the product.

Abstract: According to the present invention, proposed is a method for collecting filtered image data, the method comprising the steps of: obtaining, by a computing apparatus, first image data and recording the first image data in an image data pool; obtaining, by the computing apparatus, second image data; performing, by the computing apparatus, an operation in a deep learning model by using the first image data and the second image data as input values, and calculating a first feature vector corresponding to the first image data and a second feature vector corresponding to the second image data; and comparing, by the computing apparatus, a similarity between the first feature vector and the second feature vector, and when the similarity is less than a certain value, recording the second image data in the image data pool.