Stress Detection in It Professionals by Image Processing and Machine Learning

Abstract

The project aims to leverage image processing and machine learning techniques for the detection of stress in IT profes-sionals. The primary focus is on monitoring the emotional well-being of individuals who spend extended periods working in front of a computer, with the goal of identifying and alleviating stress to create a more conducive work environment. Key objectives include predicting stress based on observed symptoms, assessing stress levels in employees, and delivering perti-nent information. Previous research in stress detection for IT professionals has employed machine learning and image processing, prioritizing secure monitoring through captured images of authenticated users. These systems analyze stress levels using standard conversion and image processing methods, incorporating live detection and periodic analysis. In con-trast to earlier systems, they provide personalized counseling and solutions for managing both physical and mental stress levels, integrating surveys for regular assessments. Furthermore, existing systems may incorporate digital signal pro-cessing, taking into account factors such as Galvanic skin response, blood volume, pupil dilation, and skin temperature. Alternatively, they may rely on a variety of physiological signals and visual features to monitor stress levels during work.

Introduction

I. INTRODUCTION

In the current landscape, the IT industry is establishing a new benchmark in the market by introducing innovative technologies and products. This investigation sheds light on an increase in stress levels among employees within this dynamic environment. Despite the provision of mental health benefits in numerous companies, the problem remains pervasive. This research takes a closer look at the issue by aiming to recognize stress patterns among working employees in various businesses. Through the application of image processing and machine learning techniques, the objective is to assess stress patterns and identify influential factors. Machine learning algorithms, including KNN classifiers, are employed for stress categorization. The process involves capturing an employee's image through a camera as input, with initial detection performed through image processing. Image processing enhances or extracts relevant information from the image by converting it to digital form and executing operations on it. The three phases of image processing encompass importing the image using picture acquisition tools, examining and modifying the image, and generating an edited image or a report based on image analysis as the output. The Information Technology (IT) industry is renowned for its fast-paced and demanding work environment, leading to heightened stress levels among professionals. Chronic stress not only affects individual health and well-being but also results in decreased productivity, increased absenteeism, and a higher risk of burnout. Identifying and addressing stress in the IT sector is crucial for the long-term success of both employees and organizations.

This project has several key objectives, primarily centered on the development of an innovative image-based stress detection system. The system analyses facial expressions and physiological cues in photos to accurately assess stress levels. Advanced machine learning algorithms and image analysis techniques will be utilized to extract meaningful features from images, creating a robust stress assessment model. Additionally, an early warning system will be integrated to issue timely alerts to IT professionals when elevated stress levels are detected. Longitudinal data collection using images from IT professionals over time will provide insights into stress patterns and contributing factors, contributing to a deeper understanding of workplace stress.

II. LITERATURE SURVEY

Three studies focus on stress detection and analysis:

A. G. Giannakakis, D. Manousos, F. Chiarugi

1. Developed a framework for stress/anxiety detection through facial cues in video recordings. Conducted experiments inducing systematic affective states, analysing non-voluntary and semi-voluntary facial cues.
2. Investigated features such as eye-related events, mouth activity, head motion parameters, and camera-based heart rate.
3. Used feature selection and classification schemes to discriminate stress/anxiety from neutral states.
4. Proposed a ranking transformation correlating facial parameters with perceived stress/anxiety.
5. Found specific facial cues, including eye activity, mouth activity, head movements, and camera-based heart activity, as accurate indicators of stress and anxiety.

B. Nisha Raichur, Nidhi Lonakadi, Priyanka Mural

1. Addressed stress in computer-intensive work environments through real-time video analysis of facial expressions.
2. Captured non-intrusive videos and detected emotional status in each frame, making stress level decisions over sequential hours.
3. Used Theano, a Python framework, for deep learning algorithm execution, achieving positive results in predicting features.

C. U. S. Reddy, A. V. Thota, and A. Dharun

1. Examined stress disorders among IT professionals, acknowledging the increased risk due to changing lifestyles.
2. Applied machine learning techniques to analyse stress patterns using data from the OSMI mental health survey 2017.
3. Considered factors like gender, family history, and health benefits availability to identify stress influencers.
4. Boosting showed the highest accuracy among implemented models, providing insights for industries to reduce stress and create a more comfortable workplace

III. EXISTING SYSTEM

In the current framework, stress detection methods rely on digital signal processing, incorporating factors such as Galvanic skin response, blood volume, pupil dilation, and skin temperature. Another approach to addressing this issue involves considering various physiological signals and visual features, such as eye closure and head movement, for monitoring stress levels in individuals during work. However, these measurements can be intrusive and less comfortable in practical applications. Each sensor's data is compared with a stress index, which serves as a threshold value for detecting the stress level.

A. Disadvantages Of Existing System

Physiological signals employed for analysis are frequently limited by the nonstationary time performance. The extracted features explicitly provide the stress index of these physiological signals. The ECG signal is directly evaluated using the widely adopted peak J48 algorithm. Varied individuals may exhibit distinct behaviours or expressions under stress, making it challenging to identify a universal pattern for defining the stress emotion. Algorithms utilized include Bayesian Network and J48.

IV. PROPOSED SYSTEM

The proposed system utilizes machine learning algorithms, such as KNN classifiers, for stress classification. Image processing is employed in the initial stage for detection, where the employee's image is provided by the browser as input. Image processing is used to enhance the image or extract useful information by converting it into digital form and performing specific operations. The system takes an image as input, and the output may be an image or characteristics associated with that image. Emotions are displayed in a circular box, with stress levels indicated by categories such as Angry, Disgusted, Fearful, and Sad.

A. Advantages Of Proposed System

The output provides either an altered image or a report based on image analysis results. The Stress Detection System empowers employees to address their stress issues through preventive stress management solution.

Images of employees will be captured at regular intervals, and the corresponding output (stress level) will be generated in real-time, referencing the captured images.

V. SYSTEM ARCHITECTURE

VI. METHODOLOGIES

1. Image Pre-processing: The formula G(i, j) = α * F(i, j) + β is employed for image pre-processing, where α > 0 and β are gain and bias parameters, utilized to adjust the brightness and contrast of the image. Here, G(i, j) represents the output image pixel, and F(i, j) represents the input image pixel.
2. Image Resizing and Standardization: Ensuring uniform image dimensions is essential for efficient model training. Images are uniformly resized to a predefined resolution (e.g., 224x224 pixels for compatibility with common convolutional neural network architectures). Additionally, grayscale images are converted to color (if necessary) to align with the format suitable for deep learning frameworks.
3. Normalization and Pixel Value Adjustment: Normalizing pixel values across the image dataset is crucial for model convergence and stability during training. Pixel values are scaled to a range suitable for neural network models (often between 0 and 1) by dividing them by the maximum pixel value (255 for 8-bit images). Normalization helps mitigate lighting variations and enhances model robustness.

The process of feature extraction is integral to the overall pipeline, involving the transformation of raw facial expression images into detailed feature representations suitable for machine learning models. This extraction combines traditional image processing techniques with deep learning methodologies to capture distinctive and expressive features inherent in facial expressions, facilitating effective utilization in subsequent stages of the machine learning model. By employing the HOG descriptor, localized gradient distributions are computed to capture textural patterns and edge information within facial regions. The HOG descriptors quantify gradient orientations and magnitudes, encoding spatial gradients to provide insights into the textural variations characteristic of different emotional expressions.

A. Convolutional Layers

1. 1st CNN Layer: Comprising 64 filters with a kernel size of (3,3), followed by batch normalization, rectified linear unit (ReLU) activation, max-pooling, and dropout. This layer aims to extract low-level features while addressing overfitting.
2. 2nd CNN Layer: Featuring 128 filters with a kernel size of (5,5), incorporating similar layers for feature extraction and spatial reduction.
3. 3rd and 4th CNN Layers: Each composed of 512 filters, with (3,3) kernel sizes, and parallel batch normalization, activation, pooling, and dropout layers. These layers progressively learn high-level features and spatial hierarchies.

B. Fully Connected Layers

After the convolutional layers, the architecture flattens the output and connects to dense layers for higher-level feature extraction. This includes a 256-unit dense layer with batch normalization, ReLU activation, and dropout (0.25). Following this is a 512-unit dense layer with similar configurations for further abstraction and classification. The training of the Convolutional Neural Network (CNN) model follows a structured process to optimize parameters for efficient recognition and classification of emotional expressions in facial images. The initial phase involves model compilation and optimization, configuring critical parameters such as the optimizer, learning rate, loss function, and evaluation metric. The Adam optimizer, recognized for adaptive learning rates and efficient convergence, is utilized with a learning rate of 0.001. This optimizer dynamically adjusts learning rates for each parameter, enhancing the model's performance during training. Categorical cross-entropy, suitable for multi-class classification tasks, serves as the loss function to measure the disparity between predicted and actual class distributions. The primary evaluation metric is accuracy, assessing the model's correctness in predicting emotions across diverse classes.

C. Training Iterations

The model undergoes comprehensive training through the fit_generator method in Keras. This iterative training methodology efficiently manages large datasets by progressively feeding batches of pre-processed data from the training generator (train_set) into the model. The model's weights are continuously updated using backpropagation to minimize the loss function, while accuracy is enhanced through gradient descent optimization.

Conclusion

The architecture of the stress detection system emphasizes security through the automatic monitoring of authenticated user images at regular intervals, ensuring compliance with data security and privacy standards. This security-centric approach restricts image capture to authorized individuals during logged-in sessions and incorporates advanced encryption and secure data storage mechanisms. Stress detection systems in the IT sector leverage image processing, machine learning, and AI algorithms for various applica-tions, including real-time stress monitoring through facial recognition and periodic stress trend analysis. The advantages of these systems, such as early intervention, productivity improvements, and positive impacts on employee well-being, underscore their potential to transform workplace management. However, challenges, such as privacy concerns, risks of stigmatization, and the imperative for accuracy and reliability, necessitate comprehensive mitigation strategies. Future trends in stress detection systems involve emerging technologies like wearable devices and advanced AI-driven ana-lytics. Their successful integration requires adherence to ethical guidelines to ensure user privacy, data transparency, and fair usage. Achieving a balance between technological advancements and ethical considerations is essential for building trust and acceptance among employees. Organizations implementing stress detection systems are encouraged to take a holistic approach, combining technological advancements with comprehensive well-being programs. Prioritizing employee well-being within ethical frameworks and embracing evolving technologies is critical for cultivating a healthier and more produc-tive IT workforce. In summary, stress detection systems in the IT industry represent a dynamic landscape with potential benefits and challenges. Their secure design, versatile applications, and associated challenges underscore the importance of a balanced approach that prioritizes employee well-being while addressing ethical implications. Organizations should navigate this landscape thoughtfully, utilizing technological tools for employee well-being within ethical frameworks to foster a supportive and pro-ductive work environment.

References

[1] G. Giannakakis, D. Manousos, F. Chiarugi, “Stress and anxiety detection using facial cues from videos,” Biomedical Signal processing and Control”, vol. 31, pp. 89-101, January 2017. [2] T. Jick and R. Payne, “Stress at work,” Journal of Management Education, vol. 5, no. 3, pp. 50-56, 1980. [3] Nisha Raichur, Nidhi Lonakadi, Priyanka Mural, “Detection of Stress Using Image Processing and Machine Learning Techniques”, vol.9, no. 3S, July 2017. [4] Bhattacharyya, R., & Basu, S. (2018). Retrieved from ‘The Economic Times’. [5] OSMI Mental Health in Tech Survey Dataset, 2017 [6] U. S. Reddy, A. V. Thota and A. Dharun, \"Machine Learning Techniques for Stress Prediction in Working Employees,\" 2018 IEEE International Conference on Computational Intelligence and Computing Research (ICCIC), Madurai, India, 2018, pp. 1-4.

Copyright

Copyright © 2024 Nidhi Ruhil, Preet Mehta, Anushka Tyagi, Neha Sehgal, Shivendra Kumar Jha. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.