Poor-quality artificial light sources can be harmful to health and affect the productivity of workers in the workplace. This research focuses on the effects of pulse-width modulation (PWM) of artificial light sources such as displays or lamps on the human body. The main objective of the research is to make the process of checking the workplace for the presence of pulse-width modulation (PWM) more accessible and to minimize the negative impact of pulse-width modulation (PWM) on the human body. To achieve this goal, an alternative method of monitoring the pulse-width modulation (PWM) of artificial light sources has been developed using artificial intelligence (AI) methods that can even be used on a smartphone. The results of this research can be applied in the field of occupational safety and health Pulse-width modulation, PWM, artificial intelligence, AI, artificial light sources, displays, lamps, Light-emitting diode, LED, organic light emitting Diodes, OLED, active-matrix organic light emitting diodes, AMOLED, complementary metal-oxide semiconductor, CMOS, rolling shutter, occupational safety and health, OSH
With the advent of industrialization, most people began to work indoors. Artificial lighting sources
are used to organize the work process on the premises. Artificial lighting sources include both
external lighting sources and displays of computers, tablets, and smartphones with which the
employee interacts [
The main indicators of the quality of artificial lighting sources are the brightness of this source and
the pulsation coefficient. Artificial lighting sources, made from cheap components, usually have low
brightness and high ripple. The presence of a high pulsation coefficient indicates that the light source
has a flicker that is invisible to the eye. This is because such artificial light sources use the
pulsewidth modulation (PWM) method to adjust the brightness [
2023 Copyright for this paper by its authors. lighting sources and place monitors for compliance with sanitary and hygienic requirements, the workplace is attested using special equipment, namely, lux meters. This equipment is not massproduced, it is relatively expensive and, as a rule, only certification commissions have it.
Given the above, it is advisable to create an alternative to the described equipment, which would be devoid of the above disadvantages and solve the problem.
Thus, the objective of the presented research is to develop an alternative method for detecting pulse-width modulation (PWM), which would be more accessible and widespread.
To solve the problem, artificial intelligence (AI) methods will be used, which will allow the
creation of a software-hardware implementation for checking artificial light sources for the presence
of pulse-width modulation (PWM). Specifically, in this research, computer vision methods are used.
The active development of artificial intelligence (AI) and its use in various fields of activity, which
allows you to perform specific tasks more effectively, speaks of the feasibility of using this method
[
This also largely determines the scientific novelty of the research – artificial intelligence (AI) methods are used in this area for the first time, in addition, this research proposes a combination of various methods. The object of the research is the problem of the application of information technologies in the field of occupational safety and health (OSH). The subject of the research is the process of analyzing artificial lighting sources and displays at workplaces to improve the certification process for sanitary and hygienic requirements and avoid negative consequences for human health.
The results obtained in the course of this research can be used in the field of occupational safety and health (OSH). This research includes the following sections: title, authors, abstract, introduction, methodology, results, discussion, conclusion, acknowledgments, and references. These sections provide a framework for presenting the research process and contributing to the existing body of knowledge in the field.
The methodology of this research includes mixed methods, such as methods using artificial intelligence (AI), and mathematical methods. 2.1.
The use of artificial intelligence (AI) methods in various fields of activity is expedient due to its
rapid development and increasing use throughout the world. Artificial intelligence (AI) methods allow
you to perform specific tasks more subtly and efficiently [
The fact is that before training the image detection model, you need to have an array of data. Since the resulting image detection model must determine the pulse-width modulation (PWM) of artificial light sources, this imposes some limitations. Firstly, there is no ready-made data set date for this task at the moment, but it can be generated by yourself. Secondly, you need to understand what kind of training data is needed.
Based on already existing articles, it is known that you can see the pulse-width modulation (PWM)
of artificial lighting sources using a camera that has a CMOS sensor. If you fix a flickering light
source with such a camera, then black stripes will be displayed in the picture. This effect is also called
a rolling shutter (RS). This is because pulse-width modulation (PWM) is a fast flicker of an artificial
light source, and the CMOS sensor reads the image line by line, this allows you to display flicker in
the pictures [
While the camera is taking a picture line by line, the artificial light source can flicker dozens of times. However, there are also some nuances here. To see the flicker more clearly, the camera shutter speed must be very fast. As practice has shown, flicker is already visible at a shutter speed of 1/1000. At shutter speeds of 1/4000, 1/8000 flicker is even more visible. However, extremely slow shutter speeds will result in incorrect exposure, and the viewfinder image will be very dark. In this case, it is necessary to increase the light sensitivity of the matrix, which leads to image distortions.
Based on this information, it was decided to use the maximum fast shutter speed and maximum sensor light sensitivity to fix low-quality artificial light sources in the laboratory. Also, since the training model needs a variety of data to get good training results, it was a nice decision to take data from the Internet, where flicker bands are fixed on the lights.
It is necessary to take into account the current limitations, the smartphone camera is not an accurate sensor, so the classification into classes is subjective, however, it shows a trend. Based on mathematical measurements, the most popular flicker frequency values of pulse-width modulation (PWM) are: 1. less than 60 hertz (some incandescent lamps) 2. 250 hertz (common frequency for some LED/ AMOLED displays) 3. 480 hertz (common frequency for some AMOLED displays) 4. ~2000 hertz (common frequency for some new AMOLED displays) 5. ~ 25000 hertz (common frequency for some new laptops
It is important to note that OLED screens flicker more often than LED. This is due to the fact that
when the current is dimmed at low brightness values, image artifacts appear [
However, it is important to keep in mind that this data only describes a trend and may vary from one artificial light source to another. In addition, the frequency may change when the brightness of the artificial light source changes. As practice shows, it is at the lowest brightness that the lowest flicker frequency of pulse-width modulation (PWM) is achieved and, accordingly, the harm to the human body increases. Some people, knowing that, for example, their LED monitor is flickering, may use a high screen brightness, which can also harm their eyes.
Answering the question of why a person does not notice this flicker, it can be noted that the human
eye averages the flicker brightness when flickering is above 60 hertz. Flickering at 250 hertz can
cause increased eye fatigue and headaches with prolonged contact and directly affect performance.
Flickering at 500 hertz can still be detected by some people, however, the effects of it are no longer so
pronounced. Flickering over 2000 hertz is already relatively safe. Here, the brightness of an artificial
light source is often directly proportional to its flicker frequency. True only if this artificial light
source uses a pulse-width modulation method (PWM) to adjust the brightness. High-quality lighting
sources use the method of direct current regulation on the diodes (DC Dimming) [
In other words, pulse width modulation (PWM) flicker can also be described as follows. If the brightness of the artificial light source is high, then the artificial light source will burn for a longer time. If the brightness of the artificial light source is low, then the artificial light source does not burn for a longer time. This can be seen in Figure 2.
Once the data has been collected, it needs to be processed. Firstly, the dataset must be divided into
at least 2 samples - training and validation. Secondly, you need to classify the images [
Figure 3 shows the risk scale of the influence of pulse-width modulation (PWM) on the human body. Approximate data on the correspondence between health risk and qualitative amount of pulsewidth modulation (PWM) are presented in Table 1.
Let's look at example classes. They clearly show how the width of the black bars changes when the brightness of the AMOLED display changes. Figure 4 shows the class of risk "acceptable". Figure 5 shows the class of risk "tolerable". Figure 6 shows the class of risk "unacceptable". Figure 7 shows the class with no risk.
In this case, we managed to collect 100 different images for training process and 10 for validation. Objectively, it's not much. Therefore, to increase the dataset size by several times, data augmentation should be applied. Augmentation is the process of mirroring images and rotating them by several degrees. It is also important to note that since object detection methods are used on images, and not image classification, to assign the above classes to the collected data set, you must manually specify labels with their names and areas on each image. This operation is shown in Figure 8.
This research uses a model RetinaNet model for object detection pulse-width modulation (PWM).
RetinaNet is a model architecture that is commonly used for object detection tasks. It consists of two
main components: a core network (for example, MobileNetV3 in this case) for extracting features
from images, and a RetinaNet head for detecting objects based on these features [
Model architecture is shown in Figure 9. \ retina_net_model_7" model consists of several layers: "mobile_net_6": This is the MobileNet layer that takes an input image and outputs feature maps with different development levels (64x64, 32x32, 16x16, 8x8 and 1x1). MobileNet is a convolutional neural network architecture with multiple layers that uses depth wise separable convolutions and ReLU activation functions. "fpn_6": This is the Feature Pyramid Network (FPN) layer, which takes feature maps from the MobileNet layer and sequentially sets feature maps at different scales (8x8, 16x16, 32x32, 4x4 and 2x2).
"multilevel_detection_generator_7": This level is responsible for generating anchor fields at different scales and corresponding sides for each map object.
"retina_net_head_6": RetinaNet top layer that accepts feature maps from the FPN layer and performs feature detection. It outputs the coordinates of the turnover limit and probability classes for objects at different scales.
MobileNetV3 includes pretraining on a large dataset called ImageNet, which comprises millions of
labeled images from thousands of different classes. Through pretraining on ImageNet, MobileNet can
learn to recognize diverse visual patterns and features, which can subsequently be transferred and
fine-tuned for other computer vision tasks, including image classification, object detection, or
semantic segmentation, using smaller and more specific datasets [
Also, MobileNetV3 includes average pooling as a method to reduce the complexity of feature
mapping. Average pooling is a type of pooling operation that calculates the average value of a set of
objects within a pooling window [
The choice of these models is due to the fact that they are free, supported by large companies, have a balance of performance and accuracy. Thus, they are ideal for an our experimental setup.
In this research, to retrain the model, the following procedure is used. First, the model is initialized with pre-trained weights on large datasets such as ImageNet. Then, the retraining phase is carried out, adjusting the model parameters on the generated specific image detection dataset.
The model training parameters used in this research are presented in Table 2.
The selection of model training parameters is often individual and depends on specific goals and datasets. The parameters in Table 2 are often chosen empirically, but there are still some guidelines that drive the current values in this research. The main limitation here is a small dataset. With a small amount of data, the neural network can quickly remember all the examples, which can lead to overfitting. Limiting the number of epochs helps control overfitting and provides more robust model training. A higher learning rate can be chosen for faster model convergence. However, this parameter must be carefully selected to avoid convergence too fast or too slow. With a small amount of data, a smaller batch size can be chosen to improve the generalizing ability of the model and avoid overfitting. To achieve adequate accuracy of class determination, a detection accuracy of about 70% or more is required.
To form the dataset materials, a smartphone with a Sony IMX 586 CMOS and Snapdragon 765G
was used. The shutter speed was set as short as possible – 1/11626, and the sensor light sensitivity
(ISO) was as high as possible – 5699 [
To conduct experiments with retraining of the MobileNetV3 architecture for detecting images with
the flickering of pulse-width modulation (PWM) of artificial light sources, Google Collaboratory was
used as a platform for developing and executing code. Google Collaboratory provides access to
graphics processing unit (GPU) computing resources, allowing you to effectively train deep learning
models [
During the experiments, data enhancement technologies were also applied to increase the selection of a sample data set and increase the generalizing power of the models. Extension techniques such as rotations, scaling, reflection, and adding a noise library have been applied using imgaug tools and features. Using this experimental setup, the results, estimated performance, and efficiency of the pretrained architecture MobileNetV3 were obtained for the task of detecting images with the flickering pulse-width modulation (PWM) of artificial light sources. 2.2.
As mentioned above, the method for determining the pulse-width modulation (PWM) of artificial light sources using artificial intelligence (AI) methods is the main one in this research. However, to implement this method, you need to use a camera, for example, on a smartphone. Using the camera imposes some restrictions. The camera perfectly shows high-frequency ripples. For example, if the ∑ ) , flicker frequency is extremely low, for example, about 30 hertz, then black bars may not form - the camera will track the flicker like a stroboscope.
This method consists in getting an image from the camera and then averaging the RGB colors. This will give the brightness of the color. This information should be added to the array over some time, such as 1 second. Compared to the first method, this method can be more accurate but still requires the use of a camera. You can then use a formula to calculate the relative deviation of the mean (the sum of the values divided by the number of values) of the data set (x) from the maximum value (max) as a percentage. An example of a mathematical formula that can represent this method is presented below where P – is the percentage of flickering, ∑ – is the total sum of average color values, n – is the total number of average color values,
– is the max value of average color value brightness.
Thus, the formula can show how much percent the light source flickers. If it turns out 100%, then this indicates that the light source does not flicker.
This method consists in getting an illumination of light from built-in hardware light sensor in luxes. However, keep in mind that installed light sensors in portable technology often have very low sensitivity and may be useless. To implement this method, you can use the principles and approaches described in section 2.2.1. Thus, it will be necessary to collect data from the built-in light sensor for some time and apply formula 1.
This section presents the results of this research related to the detection of pulse-width modulation (PWM) of artificial light sources using artificial intelligence (AI) methods, namely computer vision to prevent their harm to human health. Here is the performance data for the object definition model.
The resulting model, based on the MobileNetV3 architecture and retraining, achieved a relatively average object detection accuracy on the test dataset. The results of validation on the test selection in COCO are shown in Table 3. (1)
The model training results show that the average accuracy (AP) at various IoU values ranges from 0.703 to 0.914. This means that the model accurately detects and classifies objects in the image. However, the average accuracy for small and medium objects cannot be calculated, which may indicate difficulties in detecting such objects. In terms of recall (AR), the model achieves a value between 0.746 and 0.754 for various IoUs and maximum detections. This means that the model is able to detect most of the objects, but some may be missed.
The total validation loss is [0.8219069242477417, 0.5985177159309387, 0.003196648322045803, 0.758350133895874], indicating that the model converges and trains well enough. Given the relatively average detection accuracy of the resulting model, quantization was not applied to it.
The essence of quantization is rounding the values of the model. This allows to reduce the speed of detection, but reduce the accuracy of detection. Checking the model on random images from the public domain. Example of successful detecting tolerable class risk on LED lamp is shown in Figure 10 with 62% probability. Example of successful detecting acceptable and tolerable class risk on displays is shown in Figure 11 with 89% and 33% probability.
It is important to note that in the final implementation of this method on practical experience in the form of a native Android application, an auxiliary method for detecting pulse-width modulation (PWM) was also additionally used by determining the average pixel color of the resulting image.
Both of these methods allow you to work in real-time live mode. It takes about 0.1 second to get the result on mid-range smartphone with SoC Snapdragon 765G.
The native Android application works as an expert system (ES) and, after analyzing the artificial light source, shows recommendations that can positively affect human health and labor productivity.
Figure 12 shows an example of a native Android application that implements both of these methods and test model in real-time mode on LED lamp in room with probability 80%. This research confirms the effectiveness and applicability of the presented methodology within this field.
This section analyzes the results of this research related to the detection of pulse-width modulation (PWM) of artificial light sources using artificial intelligence (AI) methods, namely computer vision to prevent their harm to human health.
The resulting image detection model, based on the MobileNetV3 architecture and retraining, showed relatively average accuracy and speed. The results obtained can be improved in next ways:
The very first and most important thing is to increase the size of the dataset. Instead of using
square masks, masks of different shapes can be used to label the training data. All this will also entail
a revision of the training parameters. The presented main method for detecting pulse-width
modulation (PWM) of artificial light sources using artificial intelligence (AI) methods, namely
computer vision may also have some limitations. First, the method is based on the rolling shutter (RS)
effect of CMOS sensors. However, there are matrices of an older generation, namely CCD [
In addition, a wide variety of CMOS sensors in portable devices also impose a limitation on the
objectivity of the results obtained. For example, relatively new CMOS sensors, Sony IMX 586, have a
shutter speed limit of more than 1/10000, and older ones, Sony IMX 363, have a shutter speed limit of
1/4000. This can directly affect the bar width [
This section summarizes the results of this research related to the detection of pulse-width modulation (PWM) of artificial light sources using artificial intelligence (AI) methods, namely computer vision to prevent their harm to human health and concludes its effectiveness, contribution, and opportunities for further development. As a result of experiments and analysis of this research, an alternative method for determining the pulse-width modulation (PWM) of artificial light sources using artificial intelligence (AI) methods, namely computer vision, was proposed and demonstrated. The effectiveness of the image detection model based on the MobileNetV3 architecture and additional training was shown, which confirms the applicability of the model under research in real scenarios.
In addition, to improve the results obtained, additional methods were also proposed and presented for the possible determination of the pulse-width modulation (PWM) of artificial light sources.
However, despite the results achieved, some limitations and room for improvement were also found. This includes improving the speed and accuracy of model definition.
For further development of this area of research, it is possible to consider proposals for developing a method for calculating the width of the black stripes of pulse-width modulation (PWM) in an image and comparing it with the real sizes obtained at different frequencies. Since the current solution represents a qualitative characteristic, and not a quantitative one in units of measurement.
In general, the results of this research confirm the significance and potential of the proposed methods. This work contributes to the development of the field of occupational safety and health (OSH), offering alternative available methods for attesting workplaces and preventing the negative impact of low-quality lighting sources.
We extend our heartfelt gratitude to our research team members, scientific supervisors, and faculty from Odesa Technological University (ONTU) for their invaluable contributions and support, without which this research would not have been possible.
Additionally, we appreciate the authors of the cited sources for their valuable insights. We sincerely thank all those who played a crucial role in the success of this research.