Nuclear irradiation environment image feature recovery method and system
The nuclear irradiation image feature recovery model constructed through the knowledge distillation mechanism, combined with multi-scale feature extraction and attention gating network, solves the problems of instability and low efficiency in image noise processing under nuclear irradiation environment, and achieves efficient and reliable image feature recovery.
Patent Information
- Application Number
- CN202511642192.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-02-10
AI Technical Summary
Existing technologies for image noise processing under nuclear irradiation environments suffer from unstable denoising effects, low processing efficiency, and difficulty in meeting real-time monitoring requirements. Furthermore, deep learning models have long update cycles, making it difficult to quickly respond to dynamically changing nuclear scenarios.
A knowledge distillation mechanism is used to construct a feature restoration model for nuclear irradiation images. By combining a multi-scale feature extraction network, an attention gating network, and a feature restoration network with noise mask optimization and a distillation loss function, efficient denoising and scene adaptation of image features are achieved.
It achieves efficient denoising of nuclear irradiation images, ensures the interpretability and reliability of image feature recovery, reduces the complexity of interface adaptation for system expansion, and provides high-quality and reliable image data services.
Smart Images
Figure CN121504756A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and specifically to a method and system for restoring image features in a nuclear irradiation environment. Background Technology
[0002] Image noise under nuclear irradiation is a common interference factor in nuclear facility monitoring and radiation imaging. Much of this noise is mixed with scene details and easily overlooked in routine observations. Nuclear irradiation noise is significantly destructive, causing blurred image details, distorted features, and leading to misjudgments of radiation distribution and missed detections of equipment defects. In practical applications, this manifests as decreased reliability of monitoring data and biased nuclear safety assessments, posing a challenge to the accurate conduct of nuclear industry production and research activities.
[0003] In existing technologies, feature restoration processing of nuclear irradiation images is mainly achieved through manual adjustment of filtering parameters or by relying on traditional algorithms. Traditional methods require operators to possess dual expertise in nuclear radiation imaging and image processing. Furthermore, they suffer from unstable denoising effects and low processing efficiency, making it difficult to meet the timeliness requirements of real-time monitoring of nuclear facilities. This may lead to delays in the analysis of critical information, posing potential risks to nuclear safety management.
[0004] With the development of computer vision, radiation imaging, and deep learning technologies, especially the integration of attention mechanisms and end-to-end networks, innovative solutions have been provided for feature recovery of nuclear irradiation images. Deep learning-based denoising systems can accurately focus on complex noisy regions while reducing the cost of manual intervention, improving denoising efficiency and image fidelity. However, existing deep learning denoising models often require a large amount of labeled data and computational resources to train from scratch to adapt to new radiation conditions, resulting in long model update cycles and difficulty in quickly responding to dynamic changes in nuclear scenes. If the new scene dataset has limited samples, overfitting can easily occur, affecting the stability of denoising. Summary of the Invention
[0005] Purpose of the invention: The first purpose of this invention is to provide a method for restoring the features of nuclear irradiated environment images that utilizes knowledge distillation to achieve scene universality; the second purpose is to provide a system for restoring the features of nuclear irradiated environment images.
[0006] Technical solution: A method for feature recovery from nuclear irradiation environment images, comprising the following steps:
[0007] Images of nuclear irradiation scenarios are acquired, and radiation dose and noise feature annotations are performed on the images to construct a nuclear irradiation image dataset with radiation physical characteristics.
[0008] Spatial normalization, grayscale correction, and noise masking optimization are performed on the images in the nuclear irradiation image dataset to output standardized images, thus obtaining a standardized nuclear irradiation image dataset.
[0009] An initial image feature recovery model is constructed, which includes a multi-scale feature extraction network, an attention gating network for nuclear irradiation noise regions, a feature recovery network, and an output layer connected in sequence. The initial image feature recovery model is trained using a standardized nuclear irradiation image dataset to obtain a general image feature recovery model.
[0010] Based on the knowledge distillation mechanism, a target image feature restoration model with the same architecture as the initial image feature restoration model is constructed. A distillation loss function adapted to the nuclear irradiation scene is constructed between the general image feature restoration model and the target image feature restoration model to form a nuclear irradiation image feature restoration framework model. The distillation loss function includes a nuclear irradiation scene adaptation loss term and a loss term on the ability of the target image feature restoration model to inherit the general image feature restoration model.
[0011] Images of the nuclear irradiation scene to be tested are acquired, the nuclear irradiation image feature recovery framework model is input, and the adapted target image feature recovery model is called to obtain the nuclear irradiation scene image after feature recovery.
[0012] Specifically, acquiring images of nuclear irradiation scenarios includes:
[0013] Images with different radiation doses, noise densities, and imaging devices under the same scene were acquired. At the same time, reference images without radiation interference under the same scene were also acquired. Radiation parameters and noise features were labeled on the images to construct a nuclear irradiation image dataset.
[0014] Specifically, spatial normalization is to unify the image size by scaling all original nuclear irradiation images to a fixed size using bicubic interpolation; grayscale correction is to unify the brightness by performing a linear grayscale transformation on the grayscale images; and noise mask optimization is to mark high-noise regions, segment the image, and generate a binary mask.
[0015] Specifically, the loss function formula for the initial image feature recovery model is as follows:
[0016]
[0017]
[0018] In the formula: For loss function, Weighted loss for nuclear irradiation noise region, For structural similarity loss, The total number of pixels contained in the image. For noise masking, The i-th pixel value of the denoised output image of the model. Let i be the value of the i-th pixel in the reference image. This is an indicator used to measure the structural similarity between images.
[0019] Specifically, the multi-scale feature extraction network is an encoder containing four convolutional blocks, which performs downsampling operations between convolutional blocks using max pooling, and outputs features layer by layer; the attention gating network includes three sets of dynamic weighting modules, which fuse the features output by each layer of the encoder with the features of the corresponding layer of the decoder; the feature recovery network is a decoder containing three upsampling blocks, which uses upsampling blocks to progressively upsample and restore the feature map, and the output layer uses convolutional layers in conjunction with the Tanh activation function to complete the conversion from features to image and output a denoised image.
[0020] Specifically, the formula for the distillation loss function is as follows:
[0021]
[0022]
[0023]
[0024] In the formula: This is the distillation loss function, used to balance the adaptability to nuclear irradiation scenarios and the ability to transfer knowledge. To adapt the loss term to the nuclear irradiation scenario, the denoising accuracy of the model under nuclear irradiation conditions is constrained. The target feature recovery model inherits the capability loss term of the general feature recovery model. Weighting coefficients for loss adaptation in nuclear irradiation scenarios; Weighted loss for the noise region, For structural similarity loss, These are the weighting coefficients for structural similarity loss; To output consistency loss, For loss of attentional consistency, represents the weighting coefficient for attention consistency loss.
[0025] Specifically, the knowledge distillation-driven nuclear irradiation image feature recovery framework model includes: quantifying the difference in radiation noise attention weight maps of the attention gating networks of the general image feature recovery model and the target image feature recovery model, constraining the attention pattern of the target image feature recovery model to the spatial distribution of nuclear irradiation noise to align with the general image feature recovery model; and constructing the distillation loss between the output of the target image feature recovery model and the output of the general image feature recovery model.
[0026] Preferably, the above method further includes the following steps:
[0027] When adding a new nuclear irradiation scene, perform incremental training and save the target image feature recovery model weights after incremental training optimization.
[0028] The present invention also provides a system for restoring image features of nuclear irradiation environments, comprising:
[0029] Nuclear irradiation image acquisition module: used to acquire images under nuclear irradiation scenarios, label the images with radiation dose and noise features, and construct a nuclear irradiation image dataset with radiation physical characteristics; it is also used to acquire images under the nuclear irradiation scenario to be tested, input them into the nuclear irradiation image feature recovery framework model, and obtain the feature-recovered image;
[0030] Image preprocessing module: used to perform spatial normalization, grayscale correction and noise masking optimization on images in the nuclear irradiation image dataset, outputting standardized images to obtain a standardized nuclear irradiation image dataset;
[0031] The general image feature recovery model training module is used to construct the initial image feature recovery model. The initial image feature recovery model includes a multi-scale feature extraction network, an attention gating network for nuclear irradiation noise regions, a feature recovery network, and an output layer connected in sequence. The initial image feature recovery model is trained using a standardized nuclear irradiation image dataset to obtain the general image feature recovery model.
[0032] Image Feature Restoration Module: Based on the knowledge distillation mechanism, it constructs a target image feature restoration model with the same architecture as the initial image feature restoration model. It builds a distillation loss function adapted to the nuclear irradiation scene between the general image feature restoration model and the target image feature restoration model, thus forming a nuclear irradiation image feature restoration framework model. The distillation loss function includes a nuclear irradiation scene adaptation loss term and a loss term for the target image feature restoration model's ability to inherit from the general image feature restoration model.
[0033] Preferably, the above system further includes:
[0034] User Interface Module: Connected to the image feature recovery module, it provides a graphical interactive interface, receives user-uploaded nuclear irradiation images, provides radiation dose visualization, and displays multi-dimensional quantitative indicators of model performance.
[0035] Beneficial Effects: Compared with existing technologies, the significant advantages of this invention are: This invention uses noise masking to mark noisy regions in nuclear irradiation images, clearly defining the target of processing and ensuring the interpretability and reliability of image feature restoration operations; it achieves efficient denoising of nuclear irradiation environment images based on attention-gated networks; and by optimizing the model through a distillation loss function that fuses weighted MSE of noisy regions with structural similarity loss, it ensures the quality of the feature-restored (denoised) image and the model's adaptability to new scene types, thus providing high-quality and reliable image data services for nuclear radiation monitoring. The functional modules in the image feature restoration system provided by this invention work collaboratively through standardized interfaces, without a hierarchical management mechanism. Each module records feature consistency verification information, ensuring the stability and accuracy of the image feature restoration process. Data flow between modules is achieved through an image preprocessing module, reducing the interface adaptation complexity during system expansion. Attached Figure Description
[0036] Figure 1 This is a flowchart of the method in Embodiment 1 of the present invention.
[0037] Figure 2 This is the network structure diagram of the nuclear irradiation image feature recovery framework model in Embodiment 1 of the present invention.
[0038] Figure 3 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation
[0039] A preferred embodiment of the present invention will be further described below with reference to the accompanying drawings.
[0040] Example 1
[0041] Please see Figure 1 As shown, this embodiment provides a method for restoring image features of a nuclear irradiation environment, including the following steps:
[0042] Step 1: Collect images of nuclear irradiation scenarios and construct a nuclear irradiation image dataset.
[0043] In this embodiment, noisy images from common nuclear irradiation environments are collected. The collected images are then annotated with both radiation dose and noise features to construct a general nuclear irradiation image dataset, which serves as the foundation for model training. The collected nuclear irradiation images include those with different radiation doses, noise densities, and imaging equipment. The final dataset comprises a noisy image, a noise-free reference image, a noise mask, and a radiation dose quadruple. In this embodiment, the dataset is divided into a training set and a validation set in an 8:2 ratio, providing accurate and reliable image preprocessing support for different radiation monitoring scenarios.
[0044] Step 2: Preprocess the nuclear irradiation noise dataset to obtain a standardized nuclear irradiation image dataset.
[0045] Construct an image standardization processing workflow: Perform spatial normalization, grayscale correction, and noise masking optimization on the acquired nuclear irradiation image dataset to output a standardized nuclear irradiation image dataset.
[0046] Specifically, spatial normalization unifies the image size by scaling all original nuclear irradiation images to a fixed size using bicubic interpolation, which facilitates stable training; grayscale correction unifies the brightness by performing a linear grayscale transformation on the grayscale image, so that subsequent models no longer learn brightness changes but focus on noise patterns; noise mask optimization marks high-noise regions, which serves to segment the image and generate a binary mask, making the denoising effect more accurate.
[0047] Step 3: Construct an initial image feature recovery model. The initial image feature recovery model includes a multi-scale feature extraction network, an attention gating network for nuclear irradiation noise regions, a feature recovery network, and an output layer connected in sequence. The initial image feature recovery model is pre-trained using a standardized nuclear irradiation image dataset to obtain a general image feature recovery model.
[0048] Please refer to Figure 2 As shown, an initial image feature recovery model based on an attention-gated U-shaped network is constructed, including a multi-scale feature extraction network (encoder), an attention-gated network, a feature recovery network (decoder), and an output layer. Pre-training yields a general image feature recovery model with universal nuclear irradiation noise removal capabilities. The loss function formula for the initial image feature recovery model is as follows:
[0049]
[0050]
[0051] In the formula: This is the loss function for the initial image feature restoration model, used to measure the difference between the output of the image feature restoration model and a noise-free reference image. Weighted loss is applied to the noisy region, and the mean squared error of corresponding pixel values between the denoised image and the reference image is calculated. For structural similarity loss, the denoising effect is measured from the perspective of image structural similarity. The total number of pixels contained in the image. This is a noise mask (1 represents a high noise region, and 0 represents a low noise region). The i-th pixel value of the denoised output image of the model. Let i be the value of the i-th pixel in the reference image. This is an index used to measure the structural similarity between images; the closer the value is to 1, the more similar the image spacing.
[0052] In the initial image feature restoration model, the encoder uses a 4-layer convolutional block as its core architecture, and downsampling is achieved through three 2×2 max pooling operations. It outputs multi-scale features of 64→128→256→512 channels layer by layer to build a rich feature hierarchy. The attention gating consists of three dynamic weighting modules, which perform fusion operations on the features of each layer of the encoder and the corresponding layer of the decoder to accurately locate noise regions and generate attention weight maps, providing key region guidance for subsequent denoising. The decoder relies on three upsampling blocks to progressively upsample and restore the 32×32×512 feature map, and finally restores the feature map to a size of 256×256×64, laying the foundation for outputting a clear image. The output layer uses a 1×1 convolutional layer with the Tanh activation function to process the features output by the decoder, directly completing the conversion from features to image, and outputting a denoised nuclear irradiation image.
[0053] Step 4: Construct a target image feature recovery model with the same architecture as the initial image feature recovery model. Build a distillation loss function between the general image feature recovery model and the target image feature recovery model to form a knowledge distillation-driven nuclear irradiation image feature recovery framework model.
[0054] Please refer to Figure 2 As shown, a dual-model structure is constructed based on an attention-gated U-shaped network, comprising a general image feature recovery model and a target image feature recovery model. In the constructed knowledge distillation-driven nuclear irradiation noise image feature recovery model, feature alignment, attention constraints, and output distillation are employed to achieve knowledge transfer between the models. Feature alignment minimizes the distribution difference of multi-scale feature representations between the encoders of the initial and target image feature recovery models, enabling the transfer of feature extraction capabilities from the initial model to the target model and providing a unified feature representation basis for the noise perception process. Attention constraints target the attention gating modules of the two models, quantifying the difference in their attention weight maps to constrain the target model's attention pattern to the spatial distribution of nuclear irradiation noise to align with the initial model, enabling the target model to accurately capture high-noise regions and improve the targeting and effectiveness of noise localization. Output distillation, by constructing a distillation loss between the output of the target model and the output of the initial model, achieves the transfer of general denoising knowledge from the initial model to the target model.
[0055] The formula for the distillation loss function is as follows:
[0056]
[0057]
[0058]
[0059] In the formula: The distillation loss function's core function is to balance scene adaptability and knowledge transfer capability, thereby improving training performance. The scene adaptation loss constrains the model's denoising accuracy in nuclear irradiation scenarios; The target image feature restoration model inherits the capability loss of the general image feature restoration model. The weighting coefficient for scene adaptation loss is typically set between 0.5 and 0.7 to control the balance between scene adaptation and knowledge transfer. Weighted loss for the noise region, For structural similarity loss, This is the weighting coefficient for structural similarity loss, typically taken as 0.1 to 0.3; To calculate the consistency loss, the pixel error between the output of the scene-specific target image feature restoration model and the output of the general image feature restoration model is calculated. To calculate the attention consistency loss, the difference between the attention weight map of the dedicated model output and the general image feature recovery model is calculated. The weighting coefficient for attention consistency loss is typically 0.05 to 0.1.
[0060] Step 5: When adding a new nuclear irradiation scene, perform incremental training for the new noise type in the nuclear irradiation scene, and save the target image feature recovery model weights after incremental training optimization.
[0061] Perform steps one and two within the new nuclear irradiation scenario to obtain the corresponding standardized nuclear irradiation image dataset. Train the target image feature recovery model on this standardized nuclear irradiation image dataset (which includes a dataset of new noise types within the nuclear irradiation scenario). Perform incremental training for the new noise types in the nuclear irradiation scenario:
[0062] The convergence of incremental training is determined based on the denoising effect on the validation set. Training stops when the following two conditions are met: the peak signal-to-noise ratio (PSNR) of the denoised images on the validation set is improved by ≤0.3dB; the structural similarity index (SSIM) of the denoised images on the validation set is improved by ≤0.01.
[0063] After meeting the above conditions, save the incrementally optimized target image feature recovery model weights to complete the adaptation training for the new noise type.
[0064] Step 6: Acquire images of the nuclear irradiation scene to be tested, input the nuclear irradiation image feature restoration framework model, call the corresponding target image feature restoration model, and obtain the image after feature restoration.
[0065] When a new nuclear irradiation image category is added or applied to a new scenario, repeat steps two through six.
[0066] Example 2
[0067] Please see Figure 3 As shown, this embodiment provides a nuclear irradiation environment image feature restoration system corresponding to the nuclear irradiation environment image feature restoration method described in Embodiment 1, including the following modules: image acquisition module 1, image preprocessing module 2, general image feature restoration model training module 3, image feature restoration module 4, and user interface module 5.
[0068] The above modules interact and call functions through well-defined API interfaces, forming a complete nuclear irradiation image feature restoration processing chain; image acquisition module 1 is connected to image preprocessing module 2, and image preprocessing module 2 is connected to image acquisition module 1 and image feature restoration module 3 respectively; general image feature restoration model training module 3 is connected to image preprocessing module 2 and image feature restoration module 4 respectively, and image feature restoration module 4 is connected to general image feature restoration model training module 3 and user interface module 5 respectively; user interface module 5 is connected to image feature restoration module 4.
[0069] Image acquisition module 1: Used to acquire images under nuclear irradiation scenarios and construct a nuclear irradiation noise dataset; also used to acquire images under the nuclear irradiation scenario to be tested, input them into the nuclear irradiation denoising framework model, and obtain the denoised image. In this embodiment, image acquisition module 1 is equipped with a radiation-hardened high-sensitivity back-illuminated CMOS image sensor, a radiation dose rate meter, and a high-purity germanium energy spectrometer, used to acquire raw nuclear irradiation image data in the nuclear radiation monitoring scenario in real time. Image acquisition module 1 transmits data with image preprocessing module 2 through a 10 Gigabit Ethernet interface or an InfiniBand fiber optic channel.
[0070] Image preprocessing module 2: used to preprocess the nuclear irradiation noise dataset to obtain a standardized nuclear irradiation noise dataset; in this embodiment, the image preprocessing module 2 integrates an image processing algorithm library 21 and a GPU computing unit 22, receives raw image data from the image acquisition module 1, and performs spatial normalization, grayscale correction and noise mask generation operations according to preset parameters.
[0071] The general image feature recovery model training module 3 is used to construct the initial image feature recovery model. The initial image feature recovery model includes an encoder, an attention gating network, a decoder, and an output layer connected in sequence. The initial image feature recovery model is pre-trained using a standardized nuclear irradiation noise dataset to obtain the general image feature recovery model. In this embodiment, the general image feature recovery model training module 3 is configured with a high-performance GPU computing cluster, which consists of 8 NVIDIA H100 GPUs, each equipped with 80GB of HBM3 memory, totaling 640GB of GPU memory. It is interconnected via NVLink high speed, providing more than 500 TFLOPS of FP16 computing power. It also has an internal PostgreSQL-based model library for storing model weight parameters and metadata under different training stages and application scenarios. This module provides model loading and inference services through an internal API for the image feature recovery model 4 to call, and records detailed training logs such as the loss value, validation set metric, learning rate, and parameter update status of each training batch for each training round.
[0072] Image Feature Restoration Module 4: This module constructs a target image feature restoration model with the same architecture as the initial image feature restoration model. It establishes a distillation loss function between the general image feature restoration model and the target image feature restoration model, forming a knowledge distillation-driven nuclear irradiation image feature restoration framework model. In this embodiment, the noise removal module 4 receives preprocessed noisy image data from the image preprocessing module 2 and model selection instructions from the user interface module 5 via an API interface. Based on these instructions, it loads the corresponding trained image feature restoration model from the model library of the general image feature restoration model training module 3, utilizes the GPU for parallel computation, performs image feature restoration processing on the preprocessed nuclear irradiation image, and optimizes the inference pipeline to achieve a minimum single-frame image denoising latency of 20ms.
[0073] User interface module 5: Connected to image feature restoration module 4, it provides a graphical user interface, receives user-uploaded nuclear irradiation images, calls the image preprocessing module for real-time preprocessing, and provides the function of specifying the target image feature restoration model and its parameters through selection controls. In this embodiment, user interface module 5 is built based on a web front-end technology stack and back-end API services, providing a user-friendly graphical user interface for receiving user-uploaded nuclear irradiation images, calling the image preprocessing module for real-time preprocessing, and allowing users to specify the image feature restoration model and its parameters through selection controls. Simultaneously, it can receive and display images before and after feature restoration in real-time in a dual-window comparison format, simultaneously displaying quantization denoising indicators such as PSNR, SSIM, and RMSE (root mean square error), and providing functions for downloading the denoised image and generating a PDF report containing processing details.
Claims
1. A method for restoring features from images of a nuclear irradiation environment, characterized in that, Includes the following steps: Images of nuclear irradiation scenarios are acquired, and radiation dose and noise feature annotations are performed on the images to construct a nuclear irradiation image dataset with radiation physical characteristics. Spatial normalization, grayscale correction, and noise masking optimization are performed on the images in the nuclear irradiation image dataset to output standardized images, thus obtaining a standardized nuclear irradiation image dataset. An initial image feature recovery model is constructed, which includes a multi-scale feature extraction network, an attention gating network for nuclear irradiation noise regions, a feature recovery network, and an output layer connected in sequence. The initial image feature recovery model is trained using a standardized nuclear irradiation image dataset to obtain a general image feature recovery model. Based on the knowledge distillation mechanism, a target image feature restoration model with the same architecture as the initial image feature restoration model is constructed. A distillation loss function adapted to the nuclear irradiation scene is constructed between the general image feature restoration model and the target image feature restoration model to form a nuclear irradiation image feature restoration framework model. The distillation loss function includes a nuclear irradiation scene adaptation loss term and a loss term for the target image feature restoration model's ability to inherit from the general image feature restoration model. Images of the nuclear irradiation scene to be tested are acquired, the nuclear irradiation image feature recovery framework model is input, and the adapted target image feature recovery model is called to obtain the nuclear irradiation scene image after feature recovery.
2. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that, The acquisition of images under nuclear irradiation scenarios includes: Images with different radiation doses, noise densities, and imaging devices under the same scene were acquired. At the same time, reference images without radiation interference under the same scene were also acquired. Radiation parameters and noise features were labeled on the images to construct a nuclear irradiation image dataset.
3. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that: The spatial normalization is to unify the image size, and all original nuclear irradiation images are scaled to a fixed size using bicubic interpolation; the grayscale correction is to unify the brightness, and a linear grayscale transformation is performed on the grayscale image. The noise mask is optimized by marking high-noise regions, segmenting the image, and generating a binary mask.
4. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that, The loss function formula for the initial image feature recovery model is: In the formula: For loss function, Weighted loss for nuclear irradiation noise region, For structural similarity loss, The total number of pixels contained in the image. For noise masking, The i-th pixel value of the denoised output image of the model. Let i be the value of the i-th pixel in the reference image. This is an indicator used to measure the structural similarity between images.
5. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that: The multi-scale feature extraction network is an encoder containing four convolutional blocks, which performs downsampling operations between convolutional blocks using max pooling, and outputs features layer by layer. The attention gating network includes three sets of dynamic weighting modules, which fuse the features output by each layer of the encoder with the features of the corresponding layer of the decoder. The feature recovery network is a decoder containing three upsampling blocks, which uses upsampling blocks to progressively upsample and restore the feature map. The output layer uses convolutional layers in conjunction with the Tanh activation function to complete the conversion from features to image and output a denoised image.
6. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that, The formula for the distillation loss function is as follows: In the formula: This is the distillation loss function, used to balance the adaptability to nuclear irradiation scenarios and the ability to transfer knowledge. To adapt the loss term to the nuclear irradiation scenario, the denoising accuracy of the model under nuclear irradiation conditions is constrained. The target feature recovery model inherits the capability loss term of the general feature recovery model. Weighting coefficients for loss adaptation in nuclear irradiation scenarios; Weighted loss for the noise region, For structural similarity loss, These are the weighting coefficients for structural similarity loss; To output consistency loss, For attentional consistency loss, represents the weighting coefficient for attention consistency loss.
7. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that, The knowledge distillation-driven nuclear irradiation image feature restoration framework model includes: quantifying the difference in radiation noise attention weight maps of the attention gating networks of the general image feature restoration model and the target image feature restoration model; constraining the attention pattern of the target image feature restoration model to the spatial distribution of nuclear irradiation noise to align with the general image feature restoration model; and constructing the distillation loss between the output of the target image feature restoration model and the output of the general image feature restoration model.
8. The method for restoring image features of a nuclear irradiation environment according to claim 1, characterized in that, It also includes the following steps: When adding a new nuclear irradiation scene, perform incremental training and save the target image feature recovery model weights after incremental training optimization.
9. A system for restoring image features of a nuclear irradiation environment, characterized in that, include: Nuclear irradiation image acquisition module: used to acquire images under nuclear irradiation scenarios, label the images with radiation dose and noise features, and construct a nuclear irradiation image dataset with radiation physical characteristics; it is also used to acquire images under the nuclear irradiation scenario to be tested, input them into the nuclear irradiation image feature recovery framework model, and obtain the feature-recovered image; Image preprocessing module: used to perform spatial normalization, grayscale correction and noise masking optimization on images in the nuclear irradiation image dataset, outputting standardized images to obtain a standardized nuclear irradiation image dataset; The general image feature recovery model training module is used to construct the initial image feature recovery model. The initial image feature recovery model includes a multi-scale feature extraction network, an attention gating network for nuclear irradiation noise regions, a feature recovery network, and an output layer connected in sequence. The initial image feature recovery model is trained using a standardized nuclear irradiation image dataset to obtain the general image feature recovery model. Image feature restoration module: Based on the knowledge distillation mechanism, it constructs a target image feature restoration model with the same architecture as the initial image feature restoration model. It builds a distillation loss function adapted to the nuclear irradiation scene between the general image feature restoration model and the target image feature restoration model, thus forming a nuclear irradiation image feature restoration framework model. The distillation loss function includes a nuclear irradiation scene adaptation loss term and a loss term for the target image feature restoration model's ability to inherit from the general image feature restoration model.
10. The nuclear irradiation environment image feature restoration system according to claim 9, characterized in that, Also includes: User interface module: Connected to the image feature recovery module, it provides a graphical interactive interface, receives nuclear irradiation images uploaded by users, provides radiation dose visualization function, and provides multi-dimensional quantitative index display function for model performance.