Proton energy spectrum identification method based on convolutional neural network and CMOS image sensor proton-induced transient bright spots

Abstract

The invention relates to a proton energy spectrum identification method based on a convolutional neural network and a CMOS image sensor proton-induced transient bright spot. A device involved in the method is composed of a PREF proton accelerator, a rotary sample table, a sample test board, a CMOS image sensor, a field programmable gate array and an online test computer. The method comprises the following steps: firstly, carrying out a proton online irradiation experiment with 10-60MeV proton energy and an incidence angle of 0-85 degrees, collecting a transient bright spot image output by a CMOS image sensor, extracting a single proton induced bright spot through threshold segmentation, calculating a bright spot size parameter and a pixel gray sum to construct a data set, and dividing a training set and a verification set; according to the method, a Grad-CAM visual analysis decision area is combined, a hot spot size threshold is set, effective samples are screened, proton energy identification is realized through model training, meanwhile, proton fluence measurement is completed, and the method is suitable for space detection and nuclear facility monitoring and has the advantages of being high in interference resistance, efficient in feature and flexible in adaptation.

Description

Technical Field

[0001] This invention relates to a proton energy spectrum identification method based on a convolutional neural network and a CMOS image sensor for proton-induced transient bright spots. Background Technology

[0002] CMOS image sensors, with their advantages of low cost, low power consumption, and high integration, have been widely used in consumer electronics, industrial inspection, and medical imaging, and in recent years have gradually expanded into extreme radiation environments such as space and nuclear facilities. High-energy particles in CMOS image sensor imaging not only accumulate radiation damage through ionization / displacement effects, but also induce noise such as transient bright spots through single-event effects. Although such noise is traditionally regarded as interference, the particle characteristic information it contains makes it possible for CMOS image sensors to serve as lightweight radiation detectors. Related research has attempted to use CMOS image sensors to achieve high-energy photon (γ / X-ray) detection and total ionizing dose (TID) measurement. Current radiation dose monitoring relies on specialized instruments such as gas detectors, scintillation detectors, and semiconductor detectors. While these instruments offer high accuracy, they suffer from drawbacks such as large size, high power consumption, complex signal processing, and high cost, limiting their application in space missions and portable devices. CMOS image sensors, on the other hand, can provide radiation detection auxiliary information without additional hardware, leveraging standard components already integrated into the target environment (such as satellite imaging and nuclear power plant monitoring), offering significant advantages in scene adaptability. However, current research on radiation detection using CMOS image sensors remains limited to counting protons and neutrons, and has not yet achieved incident proton energy identification based on transient bright spot morphology. Proton energy is a core parameter for analyzing radiation source characteristics, assessing radiation damage effects, and optimizing radiation protection schemes; therefore, obtaining proton energy is crucial for technological breakthroughs and practical applications in the field of radiation detection. To address the aforementioned limitations, this invention aims to explore the feasibility of using CMOS image sensors for proton energy spectrum and dose detection. It extracts transient bright spots of protons under different incident conditions through online irradiation experiments using CMOS image sensors and statistically analyzes their morphological characteristics. A ConvNet model is introduced to identify the energy of the proton bright spots, and visualization techniques are used to locate the model's region of interest. Furthermore, setting a bright spot size threshold improves the identification accuracy, ultimately providing theoretical and technical support for the development of a low-cost, lightweight proton radiation monitoring system. Summary of the Invention

[0003] The purpose of this invention is to provide a proton energy spectrum identification method based on convolutional neural networks and CMOS image sensors for proton-induced transient bright spots. The device involved in this method consists of a PREF proton accelerator, a rotating sample stage, a sample test board, a CMOS image sensor, a field-programmable gate array, and an online test computer. By optimizing the dataset construction and model training strategies, it achieves accurate identification of proton transient bright spots of different energies. The core includes four key steps: dataset processing, model training, interference elimination, and threshold optimization. First, using the raw images acquired from the proton transient irradiation experiment, a 224×224 pixel slice centered on the transient bright spot was extracted as the initial sample. Non-target bright spots within the slice were removed and filled with random background pixels to construct a basic dataset containing various energies under the actual space-orbit environment. 90% of this dataset was used for model training, and 10% for validation. The ConvNet network was selected as the base model. Grad-CAM visualization analysis confirmed that the model relied on background noise rather than bright spot morphology features for decision-making. To eliminate noise interference, the dataset was reconstructed, and the background regions of all non-target bright spots were set to zero to remove noise information. Further statistical analysis revealed a positive correlation between the model's recognition accuracy and the bright spot size. The optimal size threshold was determined through comparison to optimize the model's recognition accuracy.

[0004] The present invention discloses a proton energy spectrum identification method based on proton-induced transient bright spots using a convolutional neural network and a CMOS image sensor. The device involved in this method consists of a PREF proton accelerator (1), a rotating sample stage (2), a sample test plate (3), a CMOS image sensor (4), a field-programmable gate array (5), and an online testing computer (6). The sample test plate (3) is fixed on the rotating sample stage (2), and the CMOS image sensor (4) is placed on the sample test plate (3). The sample test plate (3) is connected to the field-programmable gate array (5), and the field-programmable gate array (5) is connected to the online testing computer (6). The specific operation is carried out according to the following steps: a. Using the PREF proton accelerator (1), a proton irradiation experiment was carried out on the CMOS image sensor (4) with proton energy of 10–60 MeV and incident angle of 0°–85°, and the dark field images output by the CMOS image sensor during the irradiation process were sampled. b. Take the image containing proton transient bright spots collected in step a, set the segmentation threshold according to the dark field noise level of the CMOS image sensor sample (4), and use the threshold segmentation algorithm to extract all proton transient bright spots in the image to obtain the initial sample image dataset of a single target transient bright spot; c. Perform noise removal processing on the sample images in the initial sample image dataset of a single target transient bright spot obtained in step b. Set the gray values ​​of all background regions in the sample images except for the target transient bright spot region to zero, completely eliminate the interference of background noise on subsequent model training, and obtain a set of bright spot sample images without noise interference. d. Construct a ConvNet model containing convolutional layers, normalization layers, and fully connected layers. The convolutional layers are used to extract the morphology and grayscale features of bright spots, the normalization layers are used to standardize the input features, and the fully connected layers are used to map the features to proton energy categories. e. Divide all images in the noise-free bright spot sample image set obtained in step c into the training set and the validation set in a ratio of 9:1. The sample label is the known energy of the corresponding proton. f. Input the training set from step e into the ConvNet model, and iterate the training model until the energy classification accuracy of the validation set tends to stabilize, with energy category classification loss as the optimization objective. g. Using Grad-CAM visualization, a heatmap of the region of interest is generated by calculating the inverse gradient of the target energy category score relative to the output of the last convolutional layer of the ConvNet model. This confirms that the region of interest for the model during inference and prediction is the bright spot region. If the heatmap of the region of interest shows that the model's region of interest is the background region, it is necessary to expand the set of bright spot sample images or adjust the hyperparameters of the ConvNet model, such as the number of network layers and channels, and retrain the ConvNet model until the model can normally focus on the bright spot region during inference and prediction. h. Calculate the ConvNet model recognition accuracy corresponding to the transient bright spot pixel size in the noise-free bright spot sample image set in step c. Calculate the bright spot size corresponding to each image sample in the validation set of the noise-free bright spot sample image set in step e. Use the model trained in step f to predict the energy of all samples in the validation set. Then, set a threshold size and calculate the prediction accuracy of all samples with sizes larger than the threshold size. By changing the size of the threshold size, obtain the relationship between the prediction accuracy and the threshold size, and use this to determine the upper and lower limits of the bright spot size selection range. i. Select bright spot samples whose size is within the range determined in step g from the set of noise-free bright spot sample images, use them as the input sample set after model optimization, and retrain the model to further improve the model's recognition stability; j. For images acquired by the CMOS image sensor in the scene to be detected, execute steps b and c sequentially to obtain a noise-free single transient bright spot sample image; filter out bright spot samples whose size is within the bright spot size selection range determined in step h, and simultaneously count the proportion of samples within the bright spot size selection range to all samples, inputting them into the ConvNet model trained in step g and optimized in step i. The model outputs the proton energy of the corresponding transient bright spot, and counts the number of proton bright spots at each energy within the bright spot size selection range. Divide this number by the proportion of samples within the bright spot size selection range to all samples to obtain the number of proton bright spots detected by the CMOS image sensor at each energy.

[0005] The proton energy spectrum identification method based on proton-induced transient bright spots using convolutional neural networks and CMOS image sensors described in this invention does not require additional dedicated radiation detection hardware. It can achieve proton energy identification and fluence statistics simply by reusing existing CMOS image sensor components. It has good robustness to background noise induced by cumulative radiation damage and is applicable not only to point-like bright spot scenarios with vertical proton incidence but also to trailing bright spot scenarios with large-angle incidence. Moreover, this method still maintains high energy identification accuracy in complex situations where proton transient bright spots overlap and large bright spots are difficult to distinguish from heavy nuclear recoil particle interference bright spots. It provides technical support for lightweight design and high-reliability application of dynamic proton radiation monitoring in scenarios such as spacecraft and nuclear facilities. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the proton transient online sampling experiment of the present invention; Figure 2 This is a schematic diagram of the training dataset samples for this invention, where the original image is on the left and the target sample image is on the right. Figure 3 This is a Grad-CAM thermal image of a typical transient bright spot sample from this invention; Figure 4 This is a schematic diagram illustrating how the average accuracy and the number of samples involved in the calculation of this invention change with the size threshold. Detailed Implementation

[0007] The present invention will now be described in further detail with reference to the accompanying drawings. Example

[0008] The proton energy spectrum identification method based on proton-induced transient bright spots using convolutional neural networks and CMOS image sensors described in this invention involves a device comprising a PREF proton accelerator 1, a rotating sample stage 2, a sample test plate 3, a CMOS image sensor 4, a field-programmable gate array (FPGA) 5, and an online testing computer 6. The sample test plate 3 is fixed on the rotating sample stage 2, and the CMOS image sensor 4 is placed on the sample test plate 3. The sample test plate 3 is connected to the FPGA 5, which is connected to the online testing computer 6. The specific operation is performed according to the following steps: a. Proton irradiation experiments with 10–60 MeV proton energies and 0°–85° incident angles were carried out on CMOS image sensor 4 using PREF proton accelerator 1, and the dark field images output by CMOS image sensor 4 during the irradiation process were sampled. b. Take the image containing proton transient bright spots acquired in step a, set a segmentation threshold according to the dark field noise level of CMOS image sensor 4, and use the threshold segmentation algorithm to extract all proton transient bright spots in the image to obtain the initial sample image dataset of a single target transient bright spot. c. Perform noise removal processing on the sample images in the initial sample image dataset of a single target transient bright spot obtained in step b. Set the gray values ​​of all background regions in the sample images except for the target transient bright spot region to zero, completely eliminate the interference of background noise on subsequent model training, and obtain a set of bright spot sample images without noise interference. d. Construct a ConvNet model containing convolutional layers, normalization layers, and fully connected layers. The convolutional layers are used to extract the morphology and grayscale features of bright spots, the normalization layers are used to standardize the input features, and the fully connected layers are used to map the features to proton energy categories. e. Divide all images in the noise-free bright spot sample image set obtained in step c into the training set and the validation set in a ratio of 9:1. The sample label is the known energy of the corresponding proton. f. Input the training set from step e into the ConvNet model, and iterate the training model until the energy classification accuracy of the validation set tends to stabilize, with energy category classification loss as the optimization objective. g. Using Grad-CAM visualization, a heatmap of the region of interest is generated by calculating the inverse gradient of the target energy category score relative to the output of the last convolutional layer of the ConvNet model. This confirms that the region of interest for the model during inference and prediction is the bright spot region. If the heatmap of the region of interest shows that the model's region of interest is the background region, it is necessary to expand the set of bright spot sample images or adjust the hyperparameters of the ConvNet model, such as the number of network layers and channels, and retrain the ConvNet model until the model can normally focus on the bright spot region during inference and prediction. h. Calculate the ConvNet model recognition accuracy corresponding to the transient bright spot pixel size in the noise-free bright spot sample image set in step c. Calculate the bright spot size corresponding to each image sample in the validation set of the noise-free bright spot sample image set in step e. Use the model trained in step f to predict the energy of all samples in the validation set. Then, set a threshold size and calculate the prediction accuracy of all samples with sizes larger than the threshold size. By changing the size of the threshold size, obtain the relationship between the prediction accuracy and the threshold size, and use this to determine the upper and lower limits of the bright spot size selection range. i. Select bright spot samples whose size is within the range determined in step g from the set of noise-free bright spot sample images, use them as the input sample set after model optimization, and retrain the model to further improve the model's recognition stability; j. For images acquired by the CMOS image sensor in the scene to be detected, execute steps b and c in sequence to obtain a noise-free single transient bright spot sample image; filter out bright spot samples whose size is within the bright spot size selection range determined in step h, and simultaneously count the proportion of samples within the bright spot size selection range to all samples, input them into the ConvNet model trained in step g and optimized in step i, the model outputs the proton energy of the corresponding transient bright spot, and counts the number of proton bright spots at each energy within the bright spot size selection range, divide the number by the proportion of samples within the bright spot size selection range to all samples, which is the number of proton bright spots detected by the CMOS image sensor at each energy; See appendix Figure 1 : a. A front-illuminated CMOS image sensor 4 from CMOSIS was selected as the detection sample. Its key parameters are: resolution of 2048×2048 pixels, pixel size of 5.5μm×5.5μm, and an 8T pixel structure with a pipelined global shutter. The sample test plate 3 was fixed on the rotating sample stage 2, and then the front-illuminated CMOS image sensor 4 from CMOSIS was placed on the sample test plate 3. The sample test plate 3 was then connected to the field programmable gate array 5, and the field programmable gate array 5 was connected to the online test computer 6. The proton irradiation experiment of the front-illuminated CMOS image sensor 4 from CMOSIS was carried out using the PREF proton accelerator 1. Three proton energies of 30MeV, 40MeV, and 50MeV were selected, with an interval of 10MeV, and eight incident angles of 0°, 15°, 30°, 45°, 60°, 75°, 80°, and 85° were set at each energy. Based on the requirement to suppress overlapping bright spots, differentiated current intensity settings are adopted for energy-angle combinations of three energies (30MeV, 40MeV, and 50MeV) and eight angles (0°, 15°, 30°, 45°, 60°, 75°, 80°, and 85°). At the same time, the beam spot area is increased and an edge irradiation method is adopted. In a dark field environment throughout the process, the operation timing and clock signal of the CMOS image sensor 4 are generated by the field-programmable gate array 5 to control image acquisition. The image data is preprocessed by the field-programmable gate array 5 and then transmitted to the online test computer 6 for storage, thus completing the sampling of dark field images containing proton transient bright spots output by the CMOS image sensor 4 during the irradiation process. b. For the image containing proton transient bright spots acquired in step a, a segmentation threshold is set according to the dark field noise level of the CMOS image sensor 4 by CMOSIS. A threshold segmentation algorithm is used to extract all proton transient bright spots in the image. A 224×224 pixel slice is made centered on each transient bright spot. For other bright spots within the slice area, they are first removed and then filled with randomly sampled background pixels, ensuring that the slice contains only a single target bright spot (with the geometric center of the bright spot as the origin). This yields the initial sample image dataset of a single target transient bright spot, as shown in the appendix. Figure 2 ; c. Perform noise removal processing on the sample images in the initial sample image dataset of a single target transient bright spot obtained in step b: set the gray values ​​of all background regions in the bright spot sample images except for the target transient bright spot region to zero, completely remove the interference of background noise on subsequent model training, and obtain a set of bright spot sample images without noise interference. This set of images contains bright spot samples with three energies of 30MeV, 40MeV and 50MeV at eight angles: 0°, 15°, 30°, 45°, 60°, 75°, 80° and 85°, with numbers of 386,443, 321,200 and 355,970 respectively. d. Construct a ConvNet model containing convolutional layers, normalization layers, and fully connected layers. The convolutional layers are used to extract the morphological features of bright spots, such as size, tail shape, cluster intensity distribution, and grayscale features. The fully connected layers are used to map these features to proton energy categories of 30MeV, 40MeV, and 50MeV. e. Divide all images in the noise-free bright spot sample image set obtained in step c into the training set and the validation set in a ratio of 9:1. The sample labels are the known energies of the corresponding protons: 30MeV, 40MeV or 50MeV. f. Input the training set into the ConvNet model, optimize the energy category classification loss, and iterate the training of the model until the energy classification accuracy of the validation set tends to stabilize. g. Using Grad-CAM visualization, a heatmap of the region of interest is generated by calculating the inverse gradient of the target energy category score relative to the output of the last convolutional layer of the ConvNet model. This confirms that the ConvNet model's region of interest during inference and prediction is the bright spot region. If the heatmap shows that the model's region of interest is the background region, it is necessary to expand the bright spot sample image set or adjust the hyperparameters of the ConvNet model, such as the number of network layers and channels, and retrain the ConvNet model until it can correctly focus on the bright spot region during inference and prediction. (See Appendix) Figure 3 ; h. Calculate the ConvNet model recognition accuracy corresponding to the transient bright spot pixel size in the noise-free bright spot sample image set in step c: Calculate the pixel size of the bright spot corresponding to each image sample in the validation set, and use the model trained in step f to predict the energy of all samples in the validation set; set a threshold size, and calculate the prediction accuracy of all samples with sizes larger than the threshold size; obtain the relationship between prediction accuracy and threshold size by changing the threshold size, and determine the upper and lower limits of the bright spot size selection range accordingly, see Appendix. Figure 4 ; i. Select bright spot samples whose size is within the range determined in step h from the set of noise-free bright spot sample images, and use them as the input sample set after model optimization. Retrain the ConvNet model to further improve the model's recognition stability. j. For images acquired by the CMOS image sensor 4 with front illumination in the target scenario, such as spacecraft or nuclear facilities, perform threshold segmentation extraction in step b and noise removal in step c to obtain a noise-free single transient bright spot sample image; select bright spot samples whose size falls within the bright spot size selection range determined in step h, and simultaneously count the proportion of samples within this range to all samples; input the selected samples into the ConvNet model optimized in step i, and the model outputs the proton energy of the corresponding transient bright spot; count the number of proton bright spots at each energy within the bright spot size selection range, and divide this number by the proportion of samples within the bright spot size selection range to all samples to obtain the number of proton bright spots detected by the CMOS image sensor at each energy.

[0009] The above description is merely a specific implementation of a method for identifying proton energy of bright spots in a CMOS image sensor based on a convolutional neural network, as described in this invention. However, the scope of protection of this invention is not limited thereto. Any substitutions or additions that can be understood by those skilled in the art within the technical scope disclosed in this invention should be included within the scope of this invention.

Claims

1.A method for identifying proton energy spectrum of proton-induced transient blooming based on a convolutional neural network and a CMOS image sensor, characterized in that, The device involved in the method is composed of a PREF proton accelerator (1), a rotating sample table (2), a sample test plate (3), a CMOS image sensor (4), a field programmable gate array (5), and an online test computer (6). The sample test plate (3) is fixed on the rotating sample table (2), the CMOS image sensor (4) is placed on the sample test plate (3), the sample test plate (3) is connected with the field programmable gate array (5), the field programmable gate array (5) is connected with the online test computer (6), and the specific operation is performed according to the following steps: a. The PREF proton accelerator (1) is used to carry out a proton irradiation experiment on the CMOS image sensor (4) at a proton energy of 10-60 MeV and an incident angle of 0°-85°, and the dark field pictures output by the CMOS image sensor during irradiation are sampled; b. The image containing the proton transient bright spot collected in step a is divided according to the dark field noise level of the CMOS image sensor (4), a threshold value is set, a threshold segmentation algorithm is used to extract all the proton transient bright spots in the image, and an initial sample image dataset of a single target transient bright spot is obtained; c. The sample images in the initial sample image dataset of a single target transient bright spot obtained in step b are subjected to noise purification treatment, the gray values of all background regions except the target transient bright spot region in the sample images are set to zero, the interference of background noise on subsequent model training is completely eliminated, and a bright spot sample image set without noise interference is obtained; d. A ConvNet model containing a convolution layer, a normalization layer and a fully connected layer is constructed, wherein the convolution layer is used to extract the morphology and gray scale features of the bright spot, the normalization is used to standardize the input features, and the fully connected layer is used to map the features to the proton energy category; e. All images in the noise-free bright spot sample image set obtained in step c are divided into a training set and a validation set according to a quantity ratio of 9:1, and the sample labels are the known energies of the corresponding protons; f. The training set in step e is input into the ConvNet model, the energy category classification loss is used as the optimization objective, and the model is iteratively trained until the energy classification accuracy of the validation set tends to be stable; g. Grad-CAM visualization is adopted, the reverse gradient of the target energy category score with respect to the output of the last convolution layer of the ConvNet model is calculated, a focus region heat map is generated, and it is confirmed that the focus region of the above model during inference prediction is the bright spot region. If the focus region heat map shows that the model focuses on the background region, the bright spot sample image set needs to be expanded or the network layer number and channel number of the ConvNet model need to be adjusted, and the ConvNet model needs to be retrained until the model can normally focus on the bright spot region during inference prediction; h. The ConvNet model recognition accuracy corresponding to the size of the transient bright spot pixel in the noise-free bright spot sample image set in step c is counted, the size of the bright spot corresponding to each image sample in the validation set of the noise-free bright spot sample image set in step e is counted, and the model trained in step f is used to predict the energy of all samples in the validation set. A threshold size is set again, and the prediction accuracy of all samples with a size greater than the threshold size is counted. The relationship between the prediction accuracy and the threshold size is obtained by changing the size of the threshold size, and the upper and lower limits of the bright spot size screening range are determined accordingly; i. Selecting bright spot samples with sizes within the range determined in step g from the noise-free bright spot sample image set as the input sample set of the optimized model and retraining the model to further improve the recognition stability of the model; j. For the image collected by the CMOS image sensor in the scene to be detected, steps b and c are executed in sequence to obtain a single noise-free transient bright spot sample image. Bright spot samples with sizes within the bright spot size screening range determined in step h are selected, and the proportion of samples within the bright spot size screening range to all samples is counted. The proportion is input into the ConvNet model trained in step g and optimized in step i. The model outputs the proton energy corresponding to the transient bright spot, and the number of proton bright spots at each energy within the bright spot size screening range is counted. The number is divided by the proportion of samples within the bright spot size screening range to all samples, and the number of proton bright spots detected by the CMOS image sensor at each energy is obtained.

Images