Nuclear power plant radiation safety visual identification system and method based on data generation and attention mechanism
By adopting a system of data generation and attention mechanisms in the nuclear power plant scenario, the problem of insufficient sample size and limitations of the entire human body recognition subject is solved, efficient radiation safety visual recognition is achieved, and the generalization and classification accuracy of the model are improved. Spend.
Patent Information
- Application Number
- CN202411241814.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2025-02-07
AI Technical Summary
Radiation safety visual recognition system in nuclear power plant scenarios faces the limitations of insufficient sample size, uneven data distribution and the entire human body recognition subject, resulting in unsatisfactory overfitting and classification effects in model training.
Using a system based on data generation and attention mechanism, data augmentation network is used to increase data through video keyframe extraction, target detection, and condition generation adversarial networks, and high-weight allocation of local targets is achieved by using attention mechanism to achieve fine-grained identification and target attributes Classification.
It effectively improves the generalization and classification accuracy of the model, reduces the cost of data acquisition, avoids overfitting of model training, and adapts to the radiation safety monitoring needs in nuclear power plant scenarios.
Smart Images
Figure CN119399665A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of personnel target detection and attribute classification in a small sample scenario of a nuclear power plant, and specifically relates to a nuclear power plant radiation safety visual recognition system and method based on data generation and attention mechanism. Background Art
[0002] The radiation safety visual recognition system in the small sample scenario of nuclear power plants extracts key frames from video clips, extracts fine-grained features of target personnel images through human target detection, classifies key attributes through the attention module, and makes up for the problems of insufficient data volume and uneven category distribution in real data sets through data augmentation, thereby enhancing the robustness and generalization of the system. This method can effectively monitor personnel violations and dangerous behaviors in real time, and has obvious advantages in ensuring personnel radiation safety protection in nuclear power plant scenarios. Commonly used classification methods lack methods for handling small sample situations in nuclear power plant scenarios, and often use the entire human body as the recognition subject.
[0003] Existing methods that only use actual data for model training and visual recognition have the following main limitations:
[0004] The actual sample size of nuclear power plant scenarios is small, and overfitting is prone to occur in the model training phase, resulting in unsatisfactory classification results in subsequent practice.
[0005] In existing data sets, samples under different light and angles are unevenly distributed, making it difficult to improve the generalization of the model.
[0006] Most existing classification methods use the entire human body as the identification subject, and their limitations include:
[0007] Compared with focusing on key targets, the entire human body has irrelevant feature information that occupies computing resources.
[0008] Lack of focus makes it impossible to assign higher weights to target attributes. Summary of the invention
[0009] The purpose of the present invention is to provide a nuclear power plant radiation safety visual recognition system and method based on data generation and attention mechanism. The present invention extracts key frames of video data, detects and cuts human target features, obtains classification training source data, helps the system to efficiently process the recognition target, improves the accuracy and efficiency of classification, and solves the problem of unsatisfactory separation effect; through conditional generative adversarial network, data augmentation is performed to make up for the problem of insufficient sample size of real scenes. This method can supplement effective training data in nuclear power plant scenes, reduce data acquisition costs, and avoid overfitting of model training. For the task of human local feature recognition in classification tasks, the present invention clearly allocates focus on target areas such as the upper body, hands, and lower body of the human body through local target positioning and attention mechanism.
[0010] The technical solution of the present invention is as follows: a nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism, including a video key frame extraction and target detection module, a small sample data augmentation module and a target fine-grained attribute recognition and classification module. The video key frame extraction and target detection module extracts key frames containing personnel from the video and performs target detection and recognition in these key frames; the small sample data augmentation module generates supplementary samples through a conditional generative adversarial network to ensure balanced sample distribution; the target fine-grained attribute recognition and classification module recognizes and classifies target attributes through an attention mechanism.
[0011] The video key frame extraction and target detection module adopts the maximum difference frame method, that is, after reading the video, the frame difference between two frames is taken, all frames are sorted according to the average frame difference intensity, and several pictures with higher intensity are selected as key frames; then the neural network detection algorithm is used to detect the target, and the frame with the highest resolution of the target person is selected.
[0012] The small sample data augmentation module uses a conditional generative adversarial network to generate samples. The generator generates new samples based on conditional information such as class labels. The discriminator is responsible for judging the authenticity of the generated samples, receiving data from the video key frame extraction and target detection module, and combining the generated reliable samples with the real samples to form a supplementary sample set, which is provided to the target fine-grained attribute recognition and classification module.
[0013] The target fine-grained attribute recognition and classification module uses the attention mechanism to capture key feature information and assigns high weights to key targets in the classification task. This module receives the expanded data set from the small sample data augmentation module and the original sample set to perform the final classification target detection and category determination.
[0014] The target fine-grained attribute recognition and classification module includes a step of securely identifying specific personnel attributes.
[0015] The security identification steps are as follows:
[0016] The video data obtained from the real video source is used as the input of the video key frame extraction and target detection module. The continuous video frames are parsed and extracted. Then, the neural network target detection algorithm is used to obtain the frame with the largest target resolution for the person trajectory formed in the key frame, and the data set of the person to be detected is obtained.
[0017] The real personnel dataset of the nuclear power plant scene obtained by the video key frame extraction and target detection module is used as the input of the small sample data augmentation module for data augmentation. According to the real pictures and labels of the nuclear power plant scene, the generator combines the label information in the latent space to generate samples and judges them through the discriminator. The generated reliable samples are used to expand the dataset. For specific classification targets, the small sample data augmentation module uses conditional generative adversarial networks to generate pseudo sample data;
[0018] The target fine-grained attribute recognition and classification module includes an attention mechanism to capture key feature information, and uses the expanded data set obtained by the small sample data augmentation module and the original sample set to perform classification target detection and category determination; first, the characteristic area of the attribute to be classified in the data set is clarified through target positioning, and then the target attribute is weighted by the attention mechanism to classify the target attribute.
[0019] A nuclear power plant radiation safety visual recognition method based on data generation and attention mechanism includes the following steps:
[0020] Step 1: Video key frame extraction and target detection;
[0021] Step 2: Small sample data augmentation;
[0022] Step 3: Target fine-grained attribute recognition and classification.
[0023] The step 1 is to extract key frames from the video stream, determine the key frames using the inter-frame difference method, and detect the target person in the key frames through the neural network detection algorithm, and select the frame with the highest resolution as the object for the next step of processing.
[0024] The key frames and target detection results obtained in step 1 are input into the small sample data augmentation module, and new samples are generated using the conditional generative adversarial network to increase the sample size in the data set and ensure a more even distribution of samples of each category.
[0025] The samples generated in step 2 are merged with the original dataset, and the attention mechanism is used to strengthen the focus on key target features during the classification process, thereby achieving the classification of target attributes.
[0026] The beneficial effects of the present invention are as follows: (1) Develop and apply relevant intelligent recognition technologies, make full use of the industrial video systems that have been deployed or will be deployed in the radiation control area of the nuclear power plant, analyze and monitor their videos in real time, timely discover radiation safety hazards on site, pay attention to unsafe behaviors of people and unsafe conditions of objects, and effectively improve the level of radiation safety prevention and control without increasing staffing. (2) The existing view parsing algorithms based on deep learning generally have the problems of high model complexity and hardware cost. The present invention proposes a lightweight algorithm model to realize automatic monitoring of personnel and items, and designs a "software and hardware integration" solution according to the characteristics of the actual scene. (3) It solves the problem of low generalization of model classification effect caused by insufficient samples in the radiation safety visual recognition system in the nuclear power plant scene, and has a good improvement in increasing the amount of training samples, quickly locating the recognition target, and optimizing the model performance. (4) Compared with the entire human body, the model focuses more on local targets such as the head and hands, which meets the specific practical needs of nuclear power plants, helps to identify whether personnel are wearing gloves, work clothes, etc., and classifies target attributes. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A structural diagram of a nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism provided by the present invention;
[0028] Figure 2 Schematic diagram of the generative adversarial network for a small sample scenario condition in a nuclear power plant. DETAILED DESCRIPTION
[0029] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0030] like Figure 1 As shown, a nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism includes a video key frame extraction and target detection module, a small sample data augmentation module and a target fine-grained attribute recognition and classification module. The video key frame extraction and target detection module extracts key frames containing personnel from the video, and performs target detection and recognition in these key frames; the small sample data augmentation module solves the problem of insufficient original sample size in the real scene of the nuclear power plant, and generates supplementary samples through a conditional generative adversarial network to ensure balanced sample distribution; the target fine-grained attribute recognition and classification module recognizes and classifies target attributes through an attention mechanism.
[0031] The video key frame extraction and target detection module adopts the maximum difference frame method, that is, after reading the video, the inter-frame difference between two frames is taken, and all frames are sorted according to the average inter-frame difference intensity. According to the data quality, the absolute threshold or relative threshold method can be used to select several pictures with higher intensity (the absolute threshold exceeds this value and is the key frame; the relative threshold is based on the statistical distribution of all frames to set the threshold) as key frames; then the neural network detection algorithm is used to detect the target, and the frame with the highest resolution of the target person is selected. The video key frame extraction and target detection module is the basic module of the entire detection and classification system. It provides data support for the subsequent small sample data augmentation module and the target fine-grained attribute recognition and classification module. The key frame and target detection results output by this module will be passed to the small sample data augmentation module as the basis for further processing.
[0032] The small sample data augmentation module uses conditional generative adversarial networks to generate samples. The generator generates new samples based on conditional information such as class labels, and the discriminator is responsible for judging the authenticity of the generated samples. This module receives data from the video key frame extraction and target detection modules, and the generated reliable samples will form a supplementary sample set together with the real samples and provide them to the target fine-grained attribute recognition and classification module.
[0033] The target fine-grained attribute recognition and classification module uses the attention mechanism to capture key feature information and assigns high weights to key targets in the classification task. This module receives the expanded data set from the small sample data augmentation module and the original sample set to perform the final classification target detection and category determination.
[0034] The specific steps for personnel attribute security identification are as follows:
[0035] 1. The video data obtained from the real video source is used as the input of the video key frame extraction and target detection module to parse and extract continuous video frames. Nuclear power plant staff generally work through remote management, the site is unmanned, and the background often has no dynamic changes. Therefore, the system uses a video frame difference detection strategy to capture dynamic changes, quickly locate the fragments where people appear, and then use the neural network target detection algorithm to obtain the frame with the largest target resolution for the trajectory of people formed in the key frame, so as to obtain a clear and clear data set of people to be detected.
[0036] Second, the real personnel dataset of the nuclear power plant scene obtained by the video key frame extraction and target detection module is used as the input of the small sample data augmentation module to perform data augmentation to make up for the problem of insufficient nuclear power plant scene dataset and uneven data distribution. Figure 2As shown in the figure, based on the real pictures and labels of the nuclear power plant scene, the generator combines the label information in the latent space to generate samples and judges them through the discriminator. The generated reliable samples are used to expand the data set. For specific classification targets, the small sample data augmentation module uses the conditional generative adversarial network to generate pseudo sample data. This strategy is mainly used to supplement the missing label types according to the actual distribution of the nuclear power plant scene data set.
[0037] 3. The target fine-grained attribute recognition and classification module of the small sample image classification system of the nuclear power plant scene provided by the invention includes an attention mechanism to capture key feature information. This module mainly uses the expanded data set obtained by the small sample data augmentation module and the original sample set to perform classification target detection and category determination. Compared with the entire human body, the model focuses more on local targets such as the head and hands. First, the characteristic area of the attribute to be classified in the data set is clarified through target positioning. After that, the weight of the target attribute is strengthened through the attention mechanism. The attention mechanism can directly reflect the key features of the local target, help identify whether the personnel are wearing gloves, work clothes, etc., and classify the target attributes.
[0038] A nuclear power plant radiation safety visual recognition method based on data generation and attention mechanism includes the following steps:
[0039] Step 1: Video key frame extraction and object detection
[0040] Extract key frames from the video stream, use the inter-frame difference method to determine the key frames, and use the neural network detection algorithm to detect the target person in the key frames, and select the frame with the highest resolution as the object for the next step of processing;
[0041] Step 2: Small Sample Data Augmentation
[0042] The key frames and target detection results obtained in step 1 are input into the small sample data augmentation module, and new samples are generated using the conditional generative adversarial network to increase the sample size in the data set and ensure a more even distribution of samples of each category.
[0043] Step 3: Target fine-grained attribute recognition and classification
[0044] The samples generated in step 2 are merged with the original dataset, and the attention mechanism is used to strengthen the focus on key target features during the classification process, thereby achieving accurate classification of target attributes (such as whether gloves are worn, whether work clothes are worn, etc.).
[0045] The implementation method of the nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism is as follows:
[0046] The system uses three key modules to achieve small sample data target recognition and attribute classification in nuclear power plant scenarios, namely video key frame extraction and target detection, small sample data augmentation, and target fine-grained attribute recognition and classification.
[0047] (1) The video key frame extraction and target detection module automatically extracts the core information from the video data, performs target detection and recognition in the key frames, and efficiently utilizes computing resources and avoids interference from invalid information. The video key frame extraction and target detection module can automatically extract the core information from the video data by taking the maximum difference frame, quickly and efficiently automatically capture the key content of the video file, and perform target detection and recognition in the key frames, avoiding the influence of a large number of invalid frames on the subsequent small sample classification. The target person forms a dynamic trajectory in the continuous video frames. The target is detected by the neural network detection algorithm and the frame with the highest resolution of the target person is selected. The extracted key frame is used for target positioning, recognition and detection. After obtaining the video key frame, the key target area in the scene image can be located by human feature recognition and the image can be cropped to obtain a to-be-processed image containing only the human body. The recognition and classification task in the nuclear power plant scene focuses on whether the personnel are wearing work clothes and protective equipment as required and whether there are dangerous behaviors. Therefore, in different classification target tasks, the cascade detection algorithm is used to achieve the focus on key local targets such as the head and hands.
[0048] (2) In order to solve the problem of insufficient original samples in real nuclear power plant scenarios, the conditional generative network is used to train and generate samples in combination with the corresponding characteristics of personnel in real nuclear power plant scenarios to supplement the label types with relatively low proportions in the sample set to ensure balanced sample distribution. The model is trained and verified using the supplementary sample set to avoid problems such as poor model training effect and easy overfitting, and to improve the model classification performance. In the small sample target classification task, the limited training data can easily cause problems such as poor model classification generalization effect. Therefore, the data augmentation strategy is used to ensure the quality of the generated samples and supplement a large number of training samples to improve the model performance. In the small sample data augmentation module, the conditional generative adversarial network model is used to generate supplementary samples to make up for the defects of "large model + small sample" and enhance the versatility and robustness of the model.
[0049] (3) The target fine-grained attribute recognition and classification module determines the category label of the detection target in different attribute classification tasks. In the past, classification tasks often regarded the human body as a whole. However, in the nuclear power plant scene safety visual recognition task, the focus of different tasks includes the head, upper body, lower body, etc. Therefore, achieving classification target positioning is crucial for the nuclear power plant scene safety visual recognition task. In the target fine-grained attribute recognition and classification module, the attention mechanism captures key feature information and helps the model understand the classification target by adjusting the attention weight. This invention solves the problem of low generalization of model classification effect due to insufficient samples in the radiation safety visual recognition system in the nuclear power plant scene, and has made great improvements in increasing the amount of training samples, quickly locating the recognition target, and optimizing the model performance.
[0050] Radiation safety management in nuclear power plants is of great significance for ensuring personnel safety, improving production efficiency and maintaining stable development of enterprises. The system uses visual analysis models to monitor and detect anomalies in the actual production environment of nuclear power plants in real time, and combines software and hardware integration technology to achieve intelligent, automated and refined safety management. With the continuous advancement of technology and the expansion of application fields, the promotion and application of this system to nuclear power plants can better improve the radiation safety management system of nuclear power plants.
Claims
1. A nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism, characterized by: It includes a video key frame extraction and target detection module, a small sample data augmentation module and a target fine-grained attribute recognition and classification module. The video key frame extraction and target detection module extracts key frames containing people from the video and performs target detection and recognition in these key frames; the small sample data augmentation module generates supplementary samples through a conditional generative adversarial network to ensure balanced sample distribution; the target fine-grained attribute recognition and classification module recognizes and classifies target attributes through an attention mechanism.
2. A nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism as claimed in claim 1, characterized in that: The video key frame extraction and target detection module adopts the maximum difference frame method, that is, after reading the video, the frame difference between two frames is taken, all frames are sorted according to the average frame difference intensity, and several pictures with higher intensity are selected as key frames; then the neural network detection algorithm is used to detect the target, and the frame with the highest resolution of the target person is selected.
3. A nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism as claimed in claim 1, characterized in that: The small sample data augmentation module uses a conditional generative adversarial network to generate samples. The generator generates new samples based on conditional information such as class labels. The discriminator is responsible for judging the authenticity of the generated samples, receiving data from the video key frame extraction and target detection module, and combining the generated reliable samples with the real samples to form a supplementary sample set, which is provided to the target fine-grained attribute recognition and classification module.
4. A nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism as claimed in claim 1, characterized in that: The target fine-grained attribute recognition and classification module uses the attention mechanism to capture key feature information and assigns high weights to key targets in the classification task. This module receives the expanded data set from the small sample data augmentation module and the original sample set to perform the final classification target detection and category determination.
5. A nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism as claimed in claim 4, characterized in that: The target fine-grained attribute recognition and classification module includes a step of securely identifying specific personnel attributes.
6. A nuclear power plant radiation safety visual recognition system based on data generation and attention mechanism as claimed in claim 5, characterized in that: The security identification steps are as follows: The video data obtained from the real video source is used as the input of the video key frame extraction and target detection module. The continuous video frames are parsed and extracted. Then, the neural network target detection algorithm is used to obtain the frame with the largest target resolution for the person trajectory formed in the key frame, and the data set of the person to be detected is obtained. The real personnel dataset of the nuclear power plant scene obtained by the video key frame extraction and target detection module is used as the input of the small sample data augmentation module for data augmentation. According to the real pictures and labels of the nuclear power plant scene, the generator combines the label information in the latent space to generate samples and judges them through the discriminator. The generated reliable samples are used to expand the dataset. For specific classification targets, the small sample data augmentation module uses conditional generative adversarial networks to generate pseudo sample data; The target fine-grained attribute recognition and classification module includes an attention mechanism to capture key feature information, and uses the expanded data set obtained by the small sample data augmentation module and the original sample set to perform classification target detection and category determination; first, the characteristic area of the attribute to be classified in the data set is clarified through target positioning, and then the target attribute is weighted by the attention mechanism to classify the target attribute.
7. A nuclear power plant radiation safety visual recognition method based on data generation and attention mechanism, characterized in that: The steps include: Step 1: Video key frame extraction and target detection; Step 2: Small sample data augmentation; Step 3: Target fine-grained attribute recognition and classification.
8. A nuclear power plant radiation safety visual recognition method based on data generation and attention mechanism as claimed in claim 7, characterized in that: The step 1 is to extract key frames from the video stream, determine the key frames using the inter-frame difference method, and detect the target person in the key frames through the neural network detection algorithm, and select the frame with the highest resolution as the object for the next step of processing.
9. A nuclear power plant radiation safety visual recognition method based on data generation and attention mechanism as claimed in claim 8, characterized in that: The step 2 is to input the key frames and target detection results obtained in step 1 into the small sample data augmentation module, and use the conditional generative adversarial network to generate new samples to increase the sample size in the data set and ensure that the distribution of samples of each category is more uniform.
10. A nuclear power plant radiation safety visual recognition method based on data generation and attention mechanism as claimed in claim 9, characterized in that: The step 3 is to merge the samples generated in step 2 with the original data set, and use the attention mechanism to strengthen the focus on key target features during the classification process, thereby achieving classification of target attributes.