In recent years, the ionizing radiation is more accessible since the promotion and application of nuclear technology, such as medical imaging and radiotherapy, industrial irradiation, nuclear energy production activities, which the radiation exposure has attracted more and more public attention. In the field of radiation dose monitoring, it can be measured by gas detector1, scintillation detectors2,3, and semiconductor detectors4 such dedicated instruments. High measurement accuracy can be achieved through these specialized radiation instruments, however, their output signals are difficult for a general to analyze and unaffordable to be used as a personal dosimeter since they are quite expensive. Today CMOS image sensors are used in an ever increasing variety of applications and indispensable in our life, ranging from moveable electronics, smartphones, medical care, security monitoring and other fields. At the same time, previous studies have shown that they are also sensitive to γ/X rays5,6,7. A typical CMOS imaging sensor is an integrated circuit with an array of pixel sensors which contains four main parts: the color filters, the pixel array, the digital controller, and the analog to digital convertor. The area sensitive to γ rays is pixel array which consists of millions of active pixel sensors responsible for capturing the intensity of the photon passing through and the intensity of a γ-ray is proportional to the amount of photons associated with it based on the principle of the photoelectric effect8,9.

Applied a CMOS image sensor for the measurement of charged particles and ionizing radiations that a variety of tentative works have been conduct. F. Wang et al.10 proposed an analysis method based on the relationship between back propagation neural network and gray value that established for measuring low energy γ-rays dose by using CMOS sensors without any X-/γ-ray converters. Q.Q. Cheng et al.11 explored the physical laws behind X-ray detection performed using CMOS sensors, their find that the histograms of the images prove there is a linear relationship between the pixel points and X-ray energy. Z.F. Yan et al.12 propose an image-processing algorithm which using a CMOS camera to extract the radioactive signal from a video containing moving objects, and enables nuclear radiation detection while the camera is in surveillance monitoring mode. H.G. Kang et al.13characterized the responses of the CMOS sensors in a smartphone to the X-rays, and their purpose is used a smartphone CMOS camera as a radiation warner for medical radiation exposure monitoring. Y.H. Johary et al.14report the feasibility of a smartphone camera sensor for radiation detecting which a detailed investigation of a well-reviewed smartphone application for radiation dosimetry that is available for popular smartphone devices under a calibration protocol that is typically used for the commercial calibration of radiation detectors, the results show that the smartphone CMOS sensor is sensitive to radiation doses as low as 10 µGy/h, with a linear dose response and an angular dependence.

Hence, there are several processing algorithms have been developed to extract radiation features from CMOS imaging sensors. A series of phone applications (i.e., apps), such as Radioactivity-Count15, Raydose16 is developed for ionizing radiation detection. However, these methods described in the literature above, extract features from a single recorded image and cause greater detection error in a lower radiation dose rate. Meanwhile, thermal noise from the semiconductor is another factor that contributes to the error. The most significant challenge is poor performance in weak radiation scenarios since the low detection efficiency of a CMOS imaging sensor as well as longer statistics time, as reported in the literature as low as sensitivity 7.35 cps/mSv/h, is not suitable for low-nuclear-radiation environment.

The aim of this work is to train radiation feature extraction algorithm capable to solve with high efficiency the radiation dose rate’s measurement problem. To achieve the before mentioned, a fitting computational tool based on a BPNN methodology was designed to extract the radioactive signal from the whole video for as longer as 3 min which is recorded by CMOS imaging sensors, and experiments were carried out to verify the feasibility of the proposed method.

Materials and methods

Experimental set-up

The pixel architecture of the CMOS sensor is based on silicon photodiode active pixel sensor technology. The structure of CMOS imaging sensor was shown in Fig. 1.

Fig. 1
figure 1

The structure of a CMOS imaging sensor.

Full size image

In the CMOS imaging sensors, the main part is sensitive to γ-rays is the photodiode, Fig. 2 was shown the space charge region, charge, electric field distribution and potential distribution of a PN junction.

Fig. 2
figure 2

The space charge region, charge, electric field distribution and potential distribution of a PN junction.

Full size image

When a photon impinges on the CMOS sensor, the PN junction of the photodiode breaks over and current occurs, the process are as follow: first, the radiation signal is converted to electron charge in the photodiode. Then, a source follower transistor which acts as a buffer located in each pixel, transfers the signal as a voltage onto the readout column bus. Finally, an analog-to-digital converter (ADC) converts the analog voltage to digital pixel values, which allows the calculation of the exposure rate, in this case, a bright blotch appears on the image, this phenomenon has been employed to detect low energy X/γ-rays for γ imaging or dosimetry. Compared to scintillator detector, with a photon-electron multiplier, the system based on CMOS image sensor can detect the γ ray directly. During the measurement, an aluminum foil thickness of 0.05 mm to blocks visible lights to the imaging sensor to make it in dark occasion.

The key part used in this experiment is a commercial CMOS imaging sensors produced by Omnivision Technologies Inc., which model is OV7725. The pixel size of the sensor is 6 μm × 6 μm, and optical size is 1/4 inches, which incorporates a 640 × 480 pixels (307200 pixels) image array, and its functional device is composed of the analog signal processor, dual 8 bit A/D and video port, digital signal processor, image scalar, timing generator. The CMOS imaging sensor was run in video mode.

Fig. 3
figure 3

(a) A commercial CMOS imaging sensor; (b) Under irradiation of 137Cs isotope source which covered with aluminum foil.

Full size image

The experiments were performed in the Key Laboratory of Neutronics and Radiation Safety at the Institute of Nuclear Energy Safety Technology (INEST). Here we start a measurement process, were we used two test sources (60Co and 137Cs) coupling with an attenuation block made of tungsten to adjust the radiation activity, which ranges from some µGy/h till to 1 Gy/h (around 1 Sv/h), the distance from sources to CMOS imaging sensors were also adjustable. Each measurement point was acquired over three minutes to ensure that the signal had sufficient time to stabilize, and radiation dose rate was measured by PTW MULTIDOS Webline coupled with a spherical ionization chamber which model is TW32002 which suitable for survey meter calibration and low level measurements. The details information about isotope sources is shown in Table 1. The source was aligned to the center of the imaging sensor face for all of the experiments. Figure 3 shows a photograph of the experimental setup.

Table 1 Information on 60Co and 137Cs radioactive sources.
Full size table

Modeling

The Back Propagation Neural Network (BPNN) was proposed and established by Rumelhart and other researchers in the early 1908s16, and after that BPNN became extensively used because of its better function approximation ability, and effective training method. Generally, a BP neural network consists of an input layer, an output layer, and several hidden layers. The structure of the three-layer BPNN is shown in Fig. 4.

Fig. 4
figure 4

Schematic of a three-layer back propagation neural network.

Full size image

It’s assume the number of input layer, hidden layer, output layer neurons of a BPNN respectively are n, R, and m. The connecting weights between the input layer and the hidden layer are \(V = (v_1 ,v_2 ,...,v_j ,...,v_R )\), with column vector \(v_j\) standing for the weight vector corresponding to the number j neutron in the hidden layer, \(b_j\) standing for the threshold value. The connecting weights between the hidden layer and the output layer are \(W = (w_1 ,w_2 ,...,w_k ,...,w_m )\), with the column vector \(w_k\) standing for the weight vector corresponding to the number k neuron in output layer, \(b_k\) standing for the threshold value. The input vector is \(X = (x_1 ,x_2 ,...,x_i ,...,x_n )^T\), the output vector of hidden layer \(Y = (y_1 ,y_2 ,...,y_j ,...,y_R )^T\), the output vector of output layer \(O = (o_1 ,o_2 ,...,o_k ,...,o_m )^T\), from which we can see that the input data is processed by connecting weights and functions from the input layer, and output data is finally achieved in the output layer; this means that the neural network is a nonlinear mapping from input to output.

The training process of the BPNN are divided the forward transfer stage of input information, and the backward transfer stage of error. Take a single neuron unit as an example, the input is \(\left[ {x_1 ,x_2 } \right]\), and weight value is \(\left[ {w_1 ,w_2 } \right]\), Offset value is \(w_0\), f is activation function, the forward transfer process of data is as follow:

$$z = w_0 x_0 + w_1 x_1 + w_2 x_2 = \sum_{i = 0}^2 {w_i x_i } = w^T x$$
(1)
$$y = f\sum_{i = 0}^2 {w_i x_i } = f(w^T x)$$
(2)

In BPNN, error back transfer is based on Delta learning rule, the error signal is transferred layer by layer from the output layer to the input layer, whose process is as follows: the error between the expected output value and the real output value is expressed like this:

$$E = \frac{1}{2}(t - y)^2$$
(3)

which y is expected output value and t is real output value.

The main purpose of BPNN is to correct the weight and minimize the error value. Delta learning rule is a general learning rule using gradient descending. The formula is as follows:

$$\Delta w = \alpha \frac{\delta E}{{\delta w}} = \alpha \frac{{\delta \frac{1}{2}\left[ {t - f(w^T x)} \right]^2 }}{\delta w} = - \alpha (t - y)f^{\prime}(w^T x)x$$
(4)

wherein, \(\alpha\) is the learning rate, with minus standing for gradient descending, thus, error E is connected with weights and threshold values that are adjusted through forward and backward transfer, until the weights meet the requirement.

Since the BPNN can realize arbitrary complex nonlinear mapping between input data and output data, serving as a good tool to solve complex problems, and the speed of a trained BP network model could be fast, which generally can meet requirements. After the network is trained, its predicative capability, also known as generalization ability, is relatively stronger, in general.

Data set and processing

The main work of this paper is extract and analyze of radiation features with an image-processing algorithm from a series of video produce by CMOS imaging sensor.

The CMOS imaging sensor was run a video recording mode at a desired frame rate, after that the recorded video was split into individual frames is processed with MATLAB R2021a. Color values of a certain pixel in an original image are (R, G, B), the gray value is calculated by10,17

$$G_r (x,y) = 0.299R + 0.587G + 0.114B$$
(5)

where \(G_r (x,y)\) the gray value of a certain point (x, y) in the image collected by the sensor, R is the red gray value of the image, G is the green gray value of the image, B is the blue gray value of the image. Figure 5 shown the data acquired by the imaging sensor, (a) is a single frame image irradiated by 137Cs isotope source that split from a video, (b) is 3D grayscale distribution of this frame, (c) is the gray value of the local bright blotch.

Fig. 5
figure 5

(a) Single frame image irradiated by 137Cs isotope source; (b) 3D grayscale distribution of this frame; (c) The gray value of the local bright blotch.

Full size image

The PN junction of the photodiode works in a reverse biased state. When there is no light, only a weak reverse current (generally less than 0.1 µΑ) can flow, which is called dark current. The photodiode is designed so that the area of the PN junction is relatively large in order to improve the light sensitivity. When there is light, the reverse current increases rapidly to tens of µΑ, which is called photocurrent, and the reverse current rises with the increase of light intensity. The change in light causes the photodiode current to change, converting the light signal into an electrical signal. The photocurrent is related to the wavelength of the incident light. When launching the application after fully covering the camera lens by an aluminum foil, the CMOS sensor’s noise level can be surveyed then determine the screening threshold.

In this work, statistical methods are used to determine the noise level. Under the condition of without radiation, it is approximately considered that the gray value of the image pixel follows the Gaussian distribution, and the image gray value falls within [E-3σ, E + 3σ] (E is the average value of the pixel gray value, and σ is the standard deviation of the image pixel) can be considered as a noise signal, that is, the discrimination threshold \(T = E + 3\sigma\), \(G_r (x,y) > T\) is a radiation event. The distribution of the gray value of the collected image is shown in Fig. 6. Finally, the appropriate threshold is used to extract the nuclear radiation information.

Fig. 6
figure 6

The distribution of the gray value of the collected image.

Full size image

It can be seen from the analysis in Fig. 6 that the discrimination threshold of the measurement system is \(T = 8\). The noise level was calculated by using Eqs. (6) and (7)

$$N = I_N + 3\sigma_N$$
(6)

where, \(I_N\) and \(\sigma_N\) represent the average and standard deviation of the total pixel intensity of the frames between the interval 3–5 s which corresponds to the frame number of 75–125.

$$I_N \approx \frac{1}{50}\sum_{k = 75}^{125} {I_k }, \sigma_N \approx \sqrt {{\frac{1}{50}\sum_{k = 75}^{125} {\left( {I_k - I_N } \right)^2 } }}$$
(7)

and the value k is the frame number and Ik is the total pixel intensity of the frame at the frame number k. After the CMOS camera captures the original image, it converts it into a gray scale image, and then scans the gray scale value \(G_r (x,y) > T\) of each pixel in the image in turn; first, the threshold T is preset to determine whether the point is thermal noise or event, if \(G_r (x,y) > T\), it is considered that the point is caused by radiating particles, and the 10 × 10 pixel area centered on the point is searched to determine whether the point is the largest point in the cluster of pixels. And if so, the point is considered to represent a radiation event. In order to ensure the accuracy of the judgment, the algorithm requires that at least two adjacent points of the maximum pixel point exceed the threshold T, then it is considered that a radiation particle event is detected. Figure 7 shows the algorithm flow of the detection system for identifying the number of radiation particle events in the image.

Fig. 7
figure 7

Radiation event screening flow chart.

Full size image

Subsequently, the frame was converted to the grey scale and the total pixel intensity of the frame was calculated. The frame exceeding the noise levels of the each red, green and blue component was extracted as a radiation event frame. This is as the input data of BPNN as shown in Fig. 8.

Fig. 8
figure 8

Histogram of a single frame which for a 588.30 µGy/h radiation.

Full size image

In order to realize the nonlinear function approximation, a BPNN was designed, in which the gray level histograms of the pixel values were set as an input date, and the corresponding irradiation dose rates were used as an output.

This computer code was developed under the MATLAB programming environment (program version is MATLAB 2021a), automates the stages of: pre-processing the information used to train and test the network, the selection of the learn rate, epochs, the training and testing stages of the network, the analysis of the performance of the trained network and the storing of information produced before, during and after training and testing stages for further analysis. The construction method of the model is described as follows:

1) Training data set and test data set. About 3 min video which acquired by CMOS imaging sensor were split into 1287 Gy value photos corresponding to a set of doses. At next stage, 5400 photos were taken from 27 groups with different dose rates as input data, and 200 were randomly selected from each group, the dose rates ranging from 2.1 µGy/h till to 588.30 µGy/h.

After that, 5400 single-frame images were converted to grayscale images according to Eqs. (5), 5000 sets of pixel spectrum were randomly selected to form the training set, and the remaining 400 sets were the test set to test the performance and generalization capabilities of the trained network.

2) Neuron network optimal. The training steps and the network error depend on the number of the neurons and the layers. By changing the number of neurons of the hidden layer, the training time and error were compared in the process of using the BPNN training method. The structure of neutron network after optimal was shown in Fig. 9 which consists of 3 hidden layers and 1 output layer. The logsig function has been selected as the transfer function of the hidden layer, and the transfer function of the output layer was a linear function Purelin. The main parameters of BPNN are learning rate is 1E-6, goal is 1E-7, and epochs are setting as 100.

Fig. 9
figure 9

Network structure after optimal.

Full size image

Based on the above network structure, performance of the BPNN is tested. The mean square error represents the expected difference between the predicted output and the target output, the lower the better. In Fig. 10, the x-coordinate represents the number of training iterations, and the y-coordinate represents the mean square error. Green in the figure shows the number of iterations of the network and the magnitude of the corresponding error when verifying the set’s best mean square error value.

Fig. 10
figure 10

The relationship between the number of iterations and MSE.

Full size image

Results and discussion

In the work described above, a group of 100 Gy level histograms were randomly selected from 30,000 images which radiation dose rate range from 7.74 ~ 608.85µGy/h to test the error after training of the four-layer network. Radiation dose rate was measured by a spherical ionization chamber and outputs values of the BPNN model discuss above are shown in Fig. 11.

Fig. 11
figure 11

Fitting results of experimental values and model output values.

Full size image

By analyzing the above data, the maximum error was about 36%, the measured value was 20.73µGy/h, and the network predicted value was 13.37µGy/h, with more than 85 sets of data predicting accuracy is better than 10%, which means that the CMOS imaging sensor with an appropriate algorithm has potential as a radiation warning device. Prediction results and errors of some data are shown in Fig. 12.

Fig. 12
figure 12

Prediction results and errors of 50 sets of data.

Full size image

Conclusions

In this paper, an improved BP neural network was developed, and its training approach was proposed between the gray level histograms and the irradiation dose rate, the network was properly trained based on an experimental input signals and its corresponding output signals. The neural network parameters were selected based on the fitting curves and error curves.

The experimental results show that the improved Elman network has less mean squared errors and better generalization result. This study demonstrates that the radiation measurement system can be used for the detection of γ-ray dose rate, but its sensitivity is low and suitable for dose monitoring in high radiation dose environment, which makes it hard to be used as a reliable radiation dosimeter. Therefore, the possible use of a CMOS image sensor is just warning the presence of radioactive materials in radiological incident environment, and cannot be used as a reliable dosimeter. This study suggests that a CMOS image sensor not an ideal candidate detector for irradiation, since the low sensitivity, and provides an innovative guidance for radiation detection.