Target state identification and judgment method based on deep learning
By combining custom models and efficient neural network structures, the problems of generalization ability and computational resource requirements of deep learning models in target motion state recognition are solved, achieving efficient and secure target state recognition and reducing development costs and data leakage risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEI JING ZHONG YAN CHUANG XIN KE JI YOU XIAN GONG SI
- Filing Date
- 2024-03-07
- Publication Date
- 2026-04-21
AI Technical Summary
Existing deep learning models suffer from limited generalization ability, high computational resource requirements, difficulty in recognition under dynamic backgrounds, real-time issues, and privacy and security problems in target motion state recognition. Furthermore, their reliance on large amounts of labeled data leads to high development costs.
It employs custom model training, unsupervised or semi-supervised learning, transfer learning, machine self-learning algorithms, and differential privacy technology, combined with OpenCV and efficient neural network structures, to achieve automated data collection and filtering, reduce computing resource requirements, and improve model generalization ability and security.
It reduces costs and workload, improves the accuracy and real-time performance of identification, ensures data security, reduces reliance on labeled data, and enhances the model's adaptability to different environments.
Smart Images
Figure CN121902186A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer algorithm technology, specifically a method for identifying and judging target states based on deep learning. Background Technology
[0002] The rapid development of deep learning, especially technologies such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), has provided powerful tools for processing image, video, and sequence data. These technologies can automatically extract useful features from large amounts of data and handle various complex patterns and data, making the recognition of target motion states more accurate and efficient. With the popularization of various intelligent devices and monitoring systems, a large amount of image and video data is generated and stored. This data provides rich training and validation resources for deep learning, enabling the realization of deep learning-based target motion state recognition methods. The emergence and development of high-performance computing devices such as GPUs and TPUs have provided powerful computing capabilities for the training and inference of deep learning models, greatly accelerating the development of deep learning-based target motion state recognition technology.
[0003] Existing technologies suffer from the following problems: limited model generalization ability, as deep learning models tend to overfit training data, leading to poor performance on unknown data; overfitting can be caused by overly simplistic content or data types; high computational resource requirements, as training and inference of deep learning models typically require high-performance GPUs or TPUs, increasing hardware costs and energy consumption; difficulty in recognition in dynamic backgrounds, as complex interactions between targets and backgrounds make target detection and motion state recognition more challenging; real-time issues, as deep learning typically requires significant computational resources and time for inference, posing a challenge in applications with high real-time requirements; privacy and security issues, as deep learning methods usually require large amounts of data for training, increasing the risk of data leakage and privacy violations; and data annotation issues, as traditional target motion state recognition methods often rely on large amounts of labeled data, increasing development costs. Summary of the Invention
[0004] The purpose of this invention is to provide a method for identifying and judging target states based on deep learning, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying and judging target states based on deep learning, the specific steps of which are as follows:
[0006] Step 1:
[0007] Use an open-source base model, collect the necessary data as needed, train your custom model using OpenCV (computer vision technology), and deploy it to a production environment where you can collect update and iteration information.
[0008] Step Two:
[0009] Automated technology is used to collect and filter information about real-time events.
[0010] Step 3:
[0011] Based on machine learning and deep learning, a new model is trained to enhance its functionality, building upon the existing target motion capture model.
[0012] Step Four:
[0013] At a specific time, a prompt is returned to the user, who then manually selects whether to update and iterate the model based on the collected data. After the update, the model is automatically deployed. If the model is not updated in the short term, the data is uploaded to the cloud or backed up. The data is automatically cleaned up after the storage time expires.
[0014] Step 5:
[0015] Machine learning algorithms can be used to automate complex information collection and filtering processes. By setting specific conditions, such as leaving the robot or falling ill and becoming unconscious, the collected data can be preprocessed, reducing the workload of manual data screening.
[0016] As a preferred technical solution of the present invention, the custom model described in step one learns from images or video sequences of moving targets by training a deep neural network, automatically extracts useful features, and classifies them, thereby improving the accuracy of motion state recognition and increasing the robot's recognition capability range.
[0017] As a preferred technical solution of the present invention, the information collection and screening in step two uses unsupervised or semi-supervised learning methods. By assisting training on unlabeled data or a small amount of labeled data, the dependence on a large amount of labeled data is reduced, thereby reducing development costs and the workload of data labeling.
[0018] As a preferred technical solution of the present invention, the new model described in step three adopts transfer learning technology to improve the performance of the model in different scenarios. By enhancing the generalization ability of the model, the system can better adapt to various environments and conditions.
[0019] As a preferred technical solution of the present invention, the machine self-learning algorithm described in step five uses a more efficient neural network structure to optimize the training algorithm, and can also perform model compression, thereby reducing the computational resource requirements and improving the inference speed.
[0020] As a preferred technical solution of the present invention, the specific conditions described in step five are used to identify the specific motion state of the target in a dynamic background through the special network structure of the machine self-learning algorithm, so that the system can better handle the interaction and dynamic changes between the target and the background, and provide them to the operator when necessary, allowing the operator to make his own judgment.
[0021] As a preferred embodiment of the present invention, the machine self-learning algorithm described in step five can encrypt the data and use differential privacy technology and design a mechanism to defend against adversarial attacks to protect user privacy and improve the security of the model.
[0022] As a preferred technical solution of the present invention, the automatic cleaning method for data exceeding the storage time described in step four can effectively reduce the storage amount of outdated data, thereby providing sufficient storage space for valid data.
[0023] As a preferred technical solution of the present invention, the method of robot autonomous operation data described in step five can make a pre-judgment of specific target states, thereby reducing workload and enabling operators to process specific data more promptly.
[0024] The beneficial effects of this invention are as follows:
[0025] This invention uses the system's judgment to perform a series of actions, reducing manpower consumption and computational resource requirements, thereby lowering costs. Compared to manual judgment, machine judgment considers more possibilities and avoids errors due to reduced effort, ensuring the model's generalization ability and reducing the operator's workload. At the same time, automation technology ensures data real-time performance, and subsequent single-person operation further protects the privacy of target users, prevents data leakage, and makes the data more secure. Attached Figure Description
[0026] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] like Figure 1As shown in the figure, this embodiment of the invention provides a method for identifying and judging the target state based on deep learning, and the specific steps are as follows:
[0029] Step 1:
[0030] Use an open-source base model, collect the necessary data as needed, train your custom model using OpenCV (computer vision technology), and deploy it to a production environment where you can collect update and iteration information.
[0031] Step Two:
[0032] Automated technology is used to collect and filter information about real-time events.
[0033] Step 3:
[0034] Based on machine learning and deep learning, a new model is trained to enhance its functionality, building upon the existing target motion capture model.
[0035] Step Four:
[0036] At a specific time, a prompt is returned to the user, who then manually selects whether to update and iterate the model based on the collected data. After the update, the model is automatically deployed. If the model is not updated in the short term, the data is uploaded to the cloud or backed up. The data is automatically cleaned up after the storage time expires.
[0037] Step 5:
[0038] Machine learning algorithms can be used to automate complex information collection and filtering processes. By setting specific conditions, such as leaving the robot or falling ill and becoming unconscious, the collected data can be preprocessed, reducing the workload of manual data screening.
[0039] This invention trains the required model using OpenCV (computer vision technology), and then, through automation technology and machine deep learning, it can accurately collect target state information, thereby reducing manpower consumption, saving costs, and reducing the workload of operators. At the same time, compared with manual processing, machine judgment considers more comprehensive and accurate possibilities, thus avoiding errors in information judgment due to operator loss of work. Moreover, after application, it enables operation by a small number of people or even a single person, which can maximize privacy and prevent data leakage, making data security more secure.
[0040] In step one, the custom model learns from images or video sequences of moving targets by training a deep neural network, automatically extracts useful features, and classifies them to improve the accuracy of motion state recognition and increase the robot's recognition range.
[0041] After collecting images or video information of moving targets using an open-source base model, useful features are extracted and then trained using OpenCV (computer vision technology) to obtain a custom model, making subsequent information collection more accurate.
[0042] In step two, information collection and screening uses unsupervised or semi-supervised learning methods. By assisting training with unlabeled or a small amount of labeled data, the reliance on a large amount of labeled data is reduced, thus lowering development costs and the workload of data labeling.
[0043] By using unsupervised or semi-supervised learning methods, unlabeled or poorly labeled data can be automatically extracted and systematically annotated, thereby reducing dependence on labeled data, reducing costs and workload, and improving data processing speed.
[0044] In step three, the new model employs transfer learning techniques to improve its performance in different scenarios. By enhancing the model's generalization ability, the system can better adapt to various environments and conditions.
[0045] The application of transfer learning techniques will enable models to have stronger generalization capabilities, thereby increasing the applicability of the models and improving their adaptability in different scenarios.
[0046] In step five, the machine self-learning algorithm uses a more efficient neural network structure to optimize the training algorithm and can also compress the model to reduce the demand for computing resources and improve the inference speed.
[0047] The reduced computational resource requirements not only improve the speed of algorithm inference but also make the model more timely and effective, enabling accurate analysis based on the user's real-time status, thereby reducing the delay in target status recognition.
[0048] In step five, specific conditions are used to identify the specific motion state of the target in a dynamic background through the special network structure of the machine self-learning algorithm. This enables the system to better handle the interaction and dynamic changes between the target and the background, and to provide the information to the operator when necessary, allowing the operator to make their own judgment.
[0049] By combining special network structures and algorithms with manual judgment by the operator, the accuracy of identifying the motion state of a target in a dynamic background can be improved, thus preventing omissions or misjudgments in the identification of the target state.
[0050] In step five, the machine learning algorithm can encrypt the data and use differential privacy technology and design mechanisms to defend against adversarial attacks to protect user privacy and improve the security of the model.
[0051] Differential privacy technology, adversarial attack defense mechanisms, and data encryption ensure that user privacy is fully protected. Except for the initial model training which requires professional training, subsequent upgrades and updates can be performed by a small number of people, or even by a single person, thereby further improving data security and reducing the risk of data leakage.
[0052] The automatic cleanup method in step four, which involves data that has exceeded the storage time, can effectively reduce the amount of outdated data stored, thus providing sufficient storage space for valid data.
[0053] Storing a large amount of invalid data will occupy a lot of storage space, thereby reducing data processing efficiency and increasing the pressure on data processing. Automatic cleaning after the time limit can effectively solve this problem.
[0054] In particular, the robot's autonomous data processing method in step five can pre-judge specific target states, thereby reducing workload and enabling operators to process specific data more promptly.
[0055] The autonomous operation and processing of robots will be more convenient and faster than manual data processing. Through the initial screening by robots, operators can save time in extracting specific data from a large amount of data, thereby improving the operator's work efficiency.
[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for identifying and judging target states based on deep learning, characterized in that, The specific steps are as follows: Step 1: Use an open-source base model, collect the data needed for production as required, train your custom model using OpenCV (computer vision technology), and deploy it to a production site where you can collect update and iteration information. Step Two: Automated technology is used to collect and filter information about real-time events. Step 3: Based on machine learning and deep learning, a new model is trained to enhance its functionality, building upon the existing target motion capture model. Step Four: At a specific time, a prompt is returned to the user, who then manually selects whether to update and iterate the model based on the collected data. After the update, the model is automatically deployed. If the model is not updated in the short term, the data is uploaded to the cloud or backed up. The data is automatically cleaned up after the storage time expires. Step 5: Machine learning algorithms can be used to automate complex information collection and filtering processes. By setting specific conditions, such as leaving the robot or falling ill and becoming unconscious, the collected data can be preprocessed, reducing the workload of manual data screening.
2. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The custom model described in step one learns from images or video sequences of moving targets by training a deep neural network, automatically extracts useful features, and classifies them, thereby improving the accuracy of motion state recognition and increasing the robot's recognition capabilities.
3. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The information collection and filtering described in step two uses unsupervised or semi-supervised learning methods. By assisting training with unlabeled or a small amount of labeled data, the reliance on a large amount of labeled data is reduced, thus lowering development costs and the workload of data labeling.
4. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The new model described in step three employs transfer learning techniques to improve the model's performance in different scenarios. By enhancing the model's generalization ability, the system can better adapt to various environments and conditions.
5. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The machine self-learning algorithm described in step five uses a more efficient neural network structure to optimize the training algorithm, and can also perform model compression to reduce computing resource requirements and improve inference speed.
6. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The specific conditions described in step five are used to identify the specific motion state of the target in a dynamic background through the special network structure of the machine self-learning algorithm. This enables the system to better handle the interaction and dynamic changes between the target and the background, and to provide the information to the operator when necessary, allowing the operator to make their own judgment.
7. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The machine learning algorithm described in step five can encrypt the data and use differential privacy technology and design mechanisms to defend against adversarial attacks to protect user privacy and improve the security of the model.
8. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The automatic cleanup method described in step four, which involves data that has exceeded the storage time, can effectively reduce the amount of outdated data stored, thereby providing sufficient storage space for valid data.
9. The method for identifying and judging target states based on deep learning according to claim 1, characterized in that: The method of autonomous data processing by the robot described in step five can make advance judgments on specific target states, thereby reducing workload and enabling operators to process specific data more promptly.