An image tagging method

The CLIP-based image tagging method automates the detection of geometric regions and selects optimal image areas for AI training, addressing the inefficiencies of manual tagging and reducing time from 8 to 2 hours.

WO2026240731A1PCT designated stage Publication Date: 2026-11-19AYGAZ ANONIM SIRKETI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/TR2026/050325
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2026-03-31
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Manual image tagging for training artificial intelligence applications in industrial systems is time-consuming and costly, necessitating a more efficient and automated process.

Method used

An image tagging method utilizing a CLIP-based model to detect geometric regions, apply sum/difference functions or optimization options, and select optimal regions based on threshold values, enabling rapid and reliable image tagging for AI training datasets.

Benefits of technology

The method significantly reduces the time required for image tagging from 8 hours to 2 hours, providing a faster and cost-effective means to generate datasets for AI modeling.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present invention discloses an image tagging method for tagging objects in a video to create a dataset for training artificial intelligence applications. The method comprises the steps of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling; receiving a video or image set comprising a plurality of products to be used in artificial intelligence modeling; performing a preprocessing by a CLIP (Contrastive Language-Image Pre-training)-based model; selecting a sum / difference function or an optimization option; placing a matrix of the received sampled data onto a pixel of a received data matrix; if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, calculating a value difference between the sampled data matrix and a portion of the image matrix on which it is located, and storing the calculated value difference matrix for the respective pixel value; if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, calculating, for all possible sum / difference functions, a value variation between the sampled data matrix and the portion of the image matrix on which it is located, and storing this value variation matrix for the respective pixel value; repeating, for each pixel, the steps of calculating the value difference between the sampled data matrix and the portion of the image matrix on which it is located and storing the calculated value difference matrix for the respective pixel value, or calculating, for all possible sum / difference functions, the value variation between the sampled data matrix and the portion of the image matrix on which it is located and storing this value variation matrix for the respective pixel value; if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, selecting a local minimum or a local maximum portion of the value difference matrix stored for each pixel value; if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, selecting, in the value variation matrix stored for each pixel value, an optimal region based on a threshold value and a size value of local minimum and local maximum regions; storing the entire results above a predetermined accuracy threshold as a tagged image.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] AN IMAGE TAGGING METHOD

[0002] Technical Field

[0003] The present invention relates to an image tagging method for tagging objects in a video.

[0004] Prior Art

[0005] With the advancement of technology, artificial intelligence applications are used, particularly for monitoring operations such as production, maintenance, and filling in industrial systems by means of image processing. In such artificial intelligence-assisted image processing applications, in order to accurately analyze the objects being monitored, said application needs to be trained using a large number of images. In order for said training process to be carried out properly, the images to be used in the training must be selected accurately. For example, in a cylinder filling facility, in order to analyze cylinders transported on a conveyor by means of an artificial intelligence-assisted image processing application, cylinders of different sizes must be grouped in the training model. In order for said training model to be created correctly, information indicating that a 12 kg household cylinder and a 2.5 kg picnic cylinder are different from each other is required. For this reason, the images to be used for training the artificial intelligence application must be tagged. The tagging process referred to herein means recognizing an object in an image (e.g., a cylinder) and / or determining its boundaries (e.g., enclosing it within a frame).

[0006] In prior art applications, obtaining images to be used for training artificial intelligence is performed manually. Here, an operator analyzes a large number of images or frames of a video and ensures tagging of the object in the image or video frame. However, performing this process manually leads to high operating costs, since it is a time-consuming process. For this reason, there is a need for an application that enables the image tagging process to be carried out rapidly and automatically.

[0007] Brief Description of Invention

[0008] The present invention discloses an image tagging method for tagging objects in a video or image set to create a dataset for training artificial intelligence applications. The method comprises receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling; receiving a video comprising a plurality of products to be used in artificial intelligence modeling; selecting a sum / difference function or an optimization option; improving accuracy and performance by reducing preprocessing results through detection of basic geometric objects by means of a CLIP (Contrastive Language-Image Pre-training)-based model; placing a matrix of the received sampled data onto a pixelin reduced images of a video matrix obtained by the CLIP-based model; if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, calculating a value difference between the sampled data matrix and a portion of the image matrix on which it is located, and storing the calculated value difference matrix for the respective pixel value; if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, calculating, for all possible sum / difference functions, a value variation between the sampled data matrix and a portion of the video image matrix on which it is located, and storing this value variation matrix for the respective pixel value; repeating, for each pixel, the steps of calculating the value difference between the sampled data matrix and the portion of the image matrix on which it is located and storing the calculated value difference matrix for the respective pixel value, or calculating, for all possible sum / difference functions, the value variation between the sampled data matrix and the portion of the image matrix on which it is located and storing the value variation matrix for the respective pixel value; if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, selecting a local minimum or a local maximum portion of the value difference matrix stored for each pixel value; if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, selecting, in the value variation matrix stored for each pixel value, an optimal region based on a threshold value and a size value of local minimum and local maximum regions; performing these processes for all regions selected in the preprocessing step by CLIP; and storing regions having an accuracy above a predetermined threshold , as tagged images.

[0009] In the method developed by the present invention, tagged images to be used fortraining an artificial intelligence-assisted image processing application are obtained using a video and sampled data. Thus, tagged images are obtained both rapidly and reliably, enabling artificial intelligence modeling to be carried out more quickly.

[0010] Objective of Invention

[0011] The object of the present invention is to develop an image tagging method for tagging objects in a video or image set.

[0012] Another object of the present invention is to develop a tagging method which works quickly and practically.

[0013] Description of Invention

[0014] Particularly in artificial intelligence-assisted image processing applications used for product tracking in industrial systems, a large number of images are needed for artificial intelligence training. Manual generation of these images by an operator leads to an increase in operating costs, as it is a time-consuming process. Accordingly, an image tagging method for taggingobjects in a video has been developed by the present invention to generate a dataset for training artificial intelligence applications.

[0015] The image tagging method developed by the present invention comprises the steps of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling; receiving a video or image set comprising a plurality of products to be used in artificial intelligence modeling; detecting basic geometric regions in the image for preprocessing by a CLIP (Contrastive Language-Image Pre-training)-based model; selecting a sum / difference function or an optimization option (e.g. by a user or operator); outputting preprocessing predictions by the CLIP-based model; placing a matrix of the received sampled data onto a pixel of data matrices obtained from the preprocessing respectively; if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, calculating a value difference between the sampled data matrix and a portion of the image matrix on which it is located, and storing the calculated value difference matrix for the respective pixel value; if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, calculating, for all possible sum / difference functions, a value variation between the sampled data matrix and the portion of the image matrix on which it is located, and storing this value variation matrix for the respective pixel value; repeating, for each pixel, the steps of calculating the value difference between the sampled data matrix and the portion of the image matrix on which it is located and storing the calculated value difference matrix for the respective pixel value, or calculating, for all possible sum / difference functions, the value variation between the sampled data matrix and the portion of the image matrix on which it is located and storing this value variation matrix for the respective pixel value; if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, selecting a local minimum or a local maximum portion of the value difference matrix stored for each pixel value; if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, selecting, in the value variation matrix stored for each pixel value, an optimal region based on a threshold value and a size value of local minimum and local maximum regions; optimizing a size of the region to be selected if this region is smaller than the region selected by CLIP; performing this process for all images selected by CLIP; selecting results above a predetermined value from among these regions; and storing the selected region as a tagged image.

[0016] In an exemplary embodiment of the present invention, said image tagging method is used in a cylinder filling facility. In this context, a video of the cylinders carried on a conveyor in the cylinder filling facility is recorded by means of a camera, wherein the recorded video is used in the step of receiving a video comprising a plurality of products to be used in artificialintelligence modeling. At least one type of cylinder carried on the conveyor (e.g. 12 kg kitchen cylinder) is taken at an angle visible in the camera where the image is taken and used as the sampled data. For example, if more than one type of cylinder is to be carried on the conveyor and if artificial intelligence modeling is to be performed according to different types of cylinders, images of different types of cylinders can also be used in the sampled data. A sum / difference function or an optimization option is selected by a user or operator for tagging the images of cylinders carried on the conveyor in the received video. Here, in both selections, the matrix of the sampled data is placed on each pixel of the matrix regions of the received image selected by the preprocessing of a CLIP model and suitable regions in the image matrix are selected accordingly. Selected regions are stored as tagged images. Thus, by tagging the products in the video, a dataset is obtained for use in artificial intelligence modeling. In the studies conducted, it was concluded that the manual tagging of a video by an operator took about 8 hours, while this process was completed in 2 hours using the developed method. Therefore, it has also been observed by experimental results that the image tagging method developed by the present invention is much faster compared to manual applications.

[0017] In a preferred embodiment according to the present invention, said method comprises, prior to the step of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling, the steps of determining whether a system (e.g. a computer) to execute the method has parallel image processing cores (e.g. CLIDA core); if said system has parallel image processing cores, checking whether an optimization compilation has been performed; if an optimization compilation has been performed, activating an optimization process and proceeding, by rendering the method (e.g. software running on the computer and using the method) deterministic, to the step of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling; if said system does not have parallel image processing cores or if an optimization compilation has not been performed, proceeding, by rendering the method (e.g. software running on the computer and using the method) deterministic, to the step of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling. The optimization process mentioned here is an application that increases the speed of the method developed with artificial intelligence support. When the optimization process is activated, the first image tags obtained by the method are used to train a preliminary model. Then, this preliminary model enables the tagging process in the rest of the video to be carried out more quickly. To achieve this, said method comprises, after the step of storing the selected region as a tagged image, if the optimization process has been activated and a predetermined number (e.g. 100) of tagged images has been obtained,training a preliminary model using the obtained tagged images and using the trained preliminary model to generate other tagged images of the video; if the optimization process has not been activated, storing obtained tagged images for the entire video.

[0018] In another preferred embodiment of the present invention, said method comprises at least one step of receiving input from the user. Here, the input received from the user may comprise a threshold value used in the step of selecting, in the value variation matrix stored for each pixel value, the region that may be the optimal based on the threshold value and size value of local minimum and local maximum regions if the optimization option is selected in the step of selecting a sum / difference function or an optimization option after first being preprocessed by a CLIP model. In a preferred embodiment, the input received from the user may comprise at least one image filter information. In this embodiment, said method comprises, prior to the step of placing the matrix of received sampled data onto a pixel of the received image matrix, applying a filter selected by the user to the received sampled data and to the received video or image set. Here, apart from the filters selected by the users, at least one predetermined filter and convolution may need to be applied in order to increase the operating speed and reliability of the system. Therefore, in another preferred embodiment according to the present invention, the method comprises, prior to the step of placing the matrix of the received sampled data onto a pixel of the received image matrix, applying at least one predetermined filter and at least one convolution to the received sampled data and to the received video.

[0019] In the method developed by the present invention, tagged images to be used fortraining an artificial intelligence-assisted image processing application are obtained using a video or set of images and sampled data. Thus, tagged images are obtained both rapidly and reliably, enabling artificial intelligence modeling to be carried out more quickly.

Claims

CLAIMS1. An image tagging method for tagging objects in a video to create a dataset for training artificial intelligence applications, characterized by comprising the steps ofreceiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling;receiving a video or image set comprising a plurality of products to be used in artificial intelligence modeling;detecting basic geometric regions in the image for preprocessing by a CLIP- based model;selecting a sum / difference function or an optimization option;outputting preprocessing predictions by the CLIP-based model;placing a matrix of the received sampled data respectively onto a pixel of data matrices obtained from the preprocessing;if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, calculating a value difference between the sampled data matrix and a portion of the image matrix on which it is located, and storing the calculated value difference matrix for the respective pixel value;if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, calculating, for all possible sum / difference functions, a value variation between the sampled data matrix and the portion of the image matrix on which it is located, and storing this value variation matrix for the respective pixel value;repeating, for each pixel, the steps of calculating the value difference between the sampled data matrix and the portion of the image matrix on which it is located and storing the calculated value difference matrix for the respective pixel value, or calculating, for all possible sum / difference functions, the value variation between the sampled data matrix and the portion of the image matrix on which it is located and storing this value variation matrix for the respective pixel value;if the sum / difference function is selected in the step of selecting a sum / difference function or an optimization option, selecting a local minimum or a local maximum portion of the value difference matrix stored for each pixelvalue;if the optimization option is selected in the step of selecting a sum / difference function or an optimization option, selecting, in the value variation matrix stored for each pixel value, an optimal region based on a threshold value and a size value of local minimum and local maximum regions;optimizing a size of the region to be selected if this region is smaller than the region selected by the CLIP-based model;performing this process for all images selected by the CLIP-based model; selecting results above a predetermined value from among these regions; and storing the selected region as a tagged image.

2. Method according to claim 1, characterized by comprising: prior to the step of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling, the steps of determining whether a system to execute the method has parallel image processing; if said system has parallel image processing cores, checking whether an optimization compilation has been performed; if an optimization compilation has been performed, activating an optimization process and proceeding, by rendering the method deterministic, to the step of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling; if said system does not have parallel image processing cores or if an optimization compilation has not been performed, proceeding, by rendering the method deterministic, to the step of receiving sampled data comprising at least one image of at least one product to be used in artificial intelligence modeling.

3. Method according to claim 2, characterized by comprising: after the step of storing the selected region as a tagged image, if the optimization process has been activated and a predetermined number of tagged images has been obtained, training a preliminary model using the obtained tagged images and using the trained preliminary model to generate other tagged images of the video; if the optimization process has not been activated, storing obtained tagged images for the entire video.

4. Method according to any of the preceding claims, characterized by comprising a step of receiving at least one input from a user.

5. Method according to claim 4, characterized in that the input received from the user comprises a threshold value used in the step of selecting, in the value variation matrix stored for each pixel value, the region that may be the optimal based on the threshold value and size value of local minimum and local maximum regions if the optimizationoption is selected in the step of selecting a sum / difference function or an optimization option.

6. Method according to claim 4 or 5, characterized in that the input received from the user comprises, if required, information relating to at least one image filter.

7. Method according to claim 6, characterized by comprising, prior to the step of placing the matrix of received sampled data onto a pixel of the received image matrix, applying, if required, a filter selected by the user to the received sampled data and to the received video or image set.

8. Method according to any of the preceding claims, characterized by comprising, prior to the step of placing the matrix of the received sampled data onto a pixel of the received image matrix, applying, if required, at least one predetermined filter and at least one convolution to the received sampled data and to the received image.