Off-line data enhancement method based on AIGC target detection

By using an offline data augmentation method based on AIGC object detection, the problems of scarce training data and loss of semantic information are solved, enabling unlimited expansion and diversification of the dataset, and improving the generalization and robustness of the model.

CN118968169AInactive Publication Date: 2024-11-15HENAN BOWO POWER TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411037999.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In existing technologies, the lack of diversity and uneven distribution of training data leads to weak generalization performance and low robustness of models. Traditional data augmentation methods generate data with limited diversity and are prone to losing semantic information.

Method used

An offline data augmentation method based on AIGC object detection is adopted. By generating a dataset on its own, the original background semantic information is preserved, and target style transformation and action transformation are performed. A dataset that meets user needs is generated using a preset sample library, thereby reducing the learning cost.

Benefits of technology

It effectively increases the amount and diversity of data, enhances the generalization and robustness of the model, and generates datasets that conform to a specific distribution and have complete semantic information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968169A_ABST
    Figure CN118968169A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of artificial intelligence data enhancement, and discloses an off-line data enhancement method based on AIGC target detection, and the method comprises the steps: carrying out the image completion of a background through an AIGC technology after a target in an image is intercepted, maintaining the semantic information in the original background as much as possible, and building a data set. The method has the advantages that the basic data is processed through the AIGC technology, semantic information in the original background can be reserved as much as possible, style transformation or action transformation is conducted on the target data, the attribute diversity of the target data is increased, sample data containing specific semantic information can be created according to the intention of a user, and the user experience is improved. According to the method, the data set with the sample features conforming to the specific distribution and the complete and reasonable semantic information is generated according to the requirements, the semantic information exceeding the semantic information which does not exist in the original data set can be created, the data set is directly generated through the preset sample library, a user does not need to provide the original data set as input, and the data set can be amplified almost infinitely and not repeatedly.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence data enhancement technology, and specifically to an offline data enhancement method based on AIGC target detection. Background Art

[0002] Generative artificial intelligence AIGC is an important symbol of the transition from the artificial intelligence 1.0 era to the 2.0 era. With the development of generative networks, the diversity and quality of data that AIGC can generate have been greatly improved. And with the growth of vehicle-side chip computing power and the development of smart cockpit-related technologies, the diversity of intelligent algorithms installed in the car has also increased. In order to develop these intelligent algorithms, developers need to collect and label a large amount of data. In the process of data processing, problems such as lack of diversity in training data, small amount of training data, and uneven distribution of training data often occur, resulting in weak model generalization performance and low robustness.

[0003] However, the main problems with the above structure are:

[0004] (1) Problems such as lack of diversity in training data, small amount of training data, and uneven distribution of training data lead to problems such as weak generalization performance and low robustness of the model;

[0005] (2) Traditional data augmentation methods generate limited data diversity;

[0006] (3) Traditional data augmentation methods will lead to the loss of semantic information in the original image and are almost unable to create new semantic information. Summary of the invention

[0007] 1. Technical issues to be solved

[0008] In view of the shortcomings of the prior art, the present invention provides an offline data enhancement method based on AIGC target detection, which has the advantages of directly generating a data set through a preset sample library, and using this to amplify the data set almost infinitely and without repetition, thereby solving the problem of semantic information loss that is prone to occur in traditional data enhancement methods.

[0009] (II) Technical solution

[0010] In order to achieve the above-mentioned purpose of directly generating a data set through a preset sample library and thereby expanding the data set almost infinitely and without duplication, the present invention provides the following technical solutions:

[0011] An offline data enhancement method based on AIGC target detection, characterized by comprising the following steps:

[0012] 1) After capturing the target in the image, the background is completed using AIGC technology, and the semantic information of the original background is retained as much as possible, and a data set is established;

[0013] 2) After acquiring the target, AIGC technology is used to transform the target data by style or action to increase the diversity of target attributes;

[0014] 3) Use the target library to store independent targets and provide more new samples when performing sample synthesis;

[0015] 4) When synthesizing new samples, since only the target itself is captured (without or very little irrelevant background), an occlusion relationship between targets is created to enrich the semantic information of the sample;

[0016] 5) Through the preset prompt library, users can select prompts to generate samples for specific scenarios, so that the generated data sets are closer to user needs and closer to real scenarios;

[0017] 6) Use the concept of offline data enhancement to reduce the user's learning cost and trial-and-error cost of different deep learning frameworks;

[0018] 7) Users can view the generated data by themselves, and it is also convenient for users to adjust parameters and generate new sample data as needed.

[0019] Preferably, in step 1), the sample is photographed and acquired by an external photographing device, the target data is analyzed by AIGC, the image background is supplemented, and the definition data in the original image is retained to complete the establishment of the data set.

[0020] Preferably, in step 2) and step 3), the data is processed by AIGC technology, the target and background of the data are extracted respectively, and the target data is style transformed or action transformed to increase the attribute diversity of the target data, the relevant data is stored, and a target library is constructed to provide more new sample selections for subsequent target synthesis.

[0021] Preferably, in step 4), when synthesizing new samples, the AIGC technology only selects the target itself, which are independent of each other. After synthesis, it can create an occlusion relationship between targets, making the sample semantic information richer.

[0022] Preferably, in step 5), a prompt library is preset to provide users with a greater number of sample targets for specific scenarios, making it convenient for users to choose, so that the generated target data set is closer to user needs and closer to real scenarios.

[0023] Preferably, in step 6) and step 7), the user can view and analyze the generated data by himself, and can modify and adjust the relevant parameters through external input devices to generate more new sample data. This can be achieved through deep learning frameworks (such as tensorflow, Pytorch, etc.), generative algorithms based on deep learning (such as stable diffusion, etc.) and programming languages ​​that can apply deep learning tools, to implement offline data enhancement technology, which can automatically correct errors when synthesizing new samples, reducing the user's learning cost and trial and error cost for different deep learning frameworks.

[0024] (III) Beneficial effects

[0025] Compared with the prior art, the present invention provides an offline data enhancement method based on AIGC target detection, which has the following beneficial effects:

[0026] This is an offline data enhancement method based on AIGC target detection. By using AIGC technology to process basic data, it can retain the semantic information in the original background as much as possible. At the same time, AIGC technology performs style transformation or action transformation on target data to increase the attribute diversity of target data. It can create sample data containing specific semantic information according to user intentions, and generate a data set with sample features that conform to a specific distribution and complete and reasonable semantic information based on this demand. It can also create semantic information that is not in the original data set, directly generate a data set through a preset sample library, without the user providing the original data set as input, and can almost infinitely and non-repetitively expand the data set, effectively increase the data volume, increase data diversity, and balance the distribution between data. Further improve model accuracy, generalization, and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a processing principle diagram of the present invention;

[0028] Figure 2 It is a schematic diagram of the processing process of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] like Figure 1 to Figure 2 As shown, an offline data enhancement method based on AIGC target detection is characterized by comprising the following steps:

[0031] 1) After capturing the target in the image, the background is completed using AIGC technology, and the semantic information of the original background is retained as much as possible, and a data set is established;

[0032] 2) After acquiring the target, AIGC technology is used to transform the target data by style or action to increase the diversity of target attributes;

[0033] 3) Use the target library to store independent targets and provide more new samples when performing sample synthesis;

[0034] 4) When synthesizing new samples, since only the target itself is captured (without or very little irrelevant background), an occlusion relationship between targets is created to enrich the semantic information of the sample;

[0035] 5) Through the preset prompt library, users can select prompts to generate samples for specific scenarios, so that the generated data sets are closer to user needs and closer to real scenarios;

[0036] 6) Use the concept of offline data enhancement to reduce the user's learning cost and trial-and-error cost of different deep learning frameworks;

[0037] 7) Users can view the generated data by themselves, and it is also convenient for users to adjust parameters and generate new sample data as needed.

[0038] In step 1), the sample is photographed and acquired through an external photographing device, the target data is analyzed through AIGC, the image background is supplemented, and the definition data in the original image is retained to complete the establishment of the data set.

[0039] In step 2) and step 3), the data is processed by AIGC technology, the target and background of the data are extracted respectively, and the style or action of the target data is transformed to increase the attribute diversity of the target data. The relevant data is stored and a target library is constructed to provide more new sample options for subsequent target synthesis.

[0040] In step 4), when synthesizing new samples, the AIGC technology only selects the target itself, which is independent of each other. After synthesis, it can create an occlusion relationship between targets, making the sample semantic information richer.

[0041] In step 5), a prompt library is preset to provide users with a larger number of sample targets for specific scenarios, which is convenient for users to choose, so that the generated target data set is closer to user needs and closer to real scenarios. The preset prompt library has the following characteristics: 1) Guiding the model: The prompt provides a starting point for the model to start generating responses, helping the model understand the user's intentions and respond accordingly; 2) Constrained generation: By adding constraints to the prompt, the style, content and logic of the model generation results can be affected; 3) Guiding output: A reasonably designed prompt can guide the model to generate the output expected by the user and improve the quality and relevance of the generated text.

[0042] In step 6) and step 7), the user can view and analyze the generated data by himself, and can modify and adjust the relevant parameters through external input devices to generate more new sample data. This can be achieved through deep learning frameworks (such as tensorflow, Pytorch, etc.), generative algorithms based on deep learning (such as stable diffusion, etc.) and programming languages ​​that can apply deep learning tools. Offline data enhancement technology can be implemented, which can automatically correct errors when synthesizing new samples, reducing the user's learning cost and trial and error cost of different deep learning frameworks.

[0043] The beneficial effects of the present invention are as follows: the offline data enhancement method based on AIGC target detection can retain the semantic information in the original background as much as possible by processing the basic data using the AIGC technology. At the same time, the AIGC technology performs style transformation or action transformation on the target data to increase the attribute diversity of the target data. Sample data containing specific semantic information can be created according to the user's intention, and a data set with sample features that conform to a specific distribution and complete and reasonable semantic information can be generated based on this demand. Semantic information that is not in the original data set can be created, and a data set can be directly generated through a preset sample library without the user providing the original data set as input. The data set can be expanded almost infinitely and non-repetitively, effectively increasing the amount of data, increasing data diversity, and balancing the distribution between data. Further improve the accuracy, generalization, and robustness of the model.

[0044] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An offline data enhancement method based on AIGC target detection, characterized in that: The following steps are involved: 1) After capturing the target in the image, the background is completed using AIGC technology, and the semantic information of the original background is retained as much as possible, and a data set is established; 2) After acquiring the target, AIGC technology is used to transform the target data by style or action to increase the diversity of target attributes; 3) Use the target library to store independent targets and provide more new samples when performing sample synthesis; 4) When synthesizing new samples, since only the target itself is captured (without or very little irrelevant background), an occlusion relationship between targets is created to enrich the semantic information of the sample; 5) Through the preset prompt library, users can select prompts to generate samples for specific scenarios, so that the generated data sets are closer to user needs and closer to real scenarios; 6) Use the concept of offline data enhancement to reduce the user's learning cost and trial-and-error cost of different deep learning frameworks; 7) Users can view the generated data by themselves, and it is also convenient for users to adjust parameters and generate new sample data as needed.

2. According to the offline data enhancement method based on AIGC target detection according to claim 1, it is characterized in that: In the step 1), the sample is photographed and acquired by an external photographing device, the target data is analyzed by AIGC, the image background is supplemented, and the definition data in the original image is retained to complete the establishment of the data set.

3. According to the offline data enhancement method based on AIGC target detection according to claim 2, it is characterized in that: In step 2) and step 3), the data is processed by AIGC technology, the target and background of the data are extracted respectively, and the target data is style transformed or action transformed to increase the attribute diversity of the target data. The relevant data is stored and a target library is constructed to provide more new sample selections for subsequent target synthesis.

4. According to the offline data enhancement method based on AIGC target detection according to claim 3, it is characterized in that: In the step 4), when synthesizing new samples, the AIGC technology only selects the target itself, which is independent of each other. After synthesis, it can create an occlusion relationship between targets, making the sample semantic information richer.

5. According to the offline data enhancement method based on AIGC target detection according to claim 4, it is characterized in that: In the step 5), a prompt library is preset to provide users with more sample target numbers for specific scenarios, which is convenient for users to choose, so that the generated target data set is closer to the user's needs and closer to the real scene.

6. According to the offline data enhancement method based on AIGC target detection according to claim 5, it is characterized in that: In step 6) and step 7), the user can view and analyze the generated data by himself, and can modify and adjust the relevant parameters through external input devices to generate more new sample data. This can be achieved through deep learning frameworks (such as tensorflow, Pytorch, etc.), generative algorithms based on deep learning (such as stable diffusion, etc.) and programming languages ​​that can apply deep learning tools, realizing offline data enhancement technology, which can correct errors when synthesizing new samples by itself, reducing the user's learning cost and trial and error cost for different deep learning frameworks.