Face detection and feature extraction for classification tasks on thermal images of Covid-19 patients

Abstract

This thesis presents a methodology for diagnosing COVID-19 patients using computer vision, infrared thermography, and machine learning. The study focuses on the analysis of thermal images, which offer a non-invasive and contactless alternative to traditional imaging methods like computed tomography (CT) and radiography. The research leverages a database of thermal images from 252 patients, including both COVID-19 positive and negative cases, to explore the potential of infrared thermography in detecting respiratory diseases. The proposed methodology involves two main approaches: one using a Convolutional Neural Network (CNN) to extract features from the full thermal image, and another incorporating a face detection step to focus on facial features. Three face detection algorithms—Haar Cascades, Local Binary Patterns (LBP), and CNNs (specifically YOLOv5)—were evaluated, with achieved accuracies of 93%, 98%, and 100%, respectively. Feature extraction was performed using the VGG-16 CNN architecture, pre-trained on the ImageNet dataset, followed by classification using traditional machine learning models such as Logistic Regression, AdaBoost, Support Vector Machines (SVM), Random Forest, and Gradient Boosting. The methodology was tested on two classification tasks: gender classification and COVID-19 symptom classification. For gender classification, the full-body approach achieved accuracies ranging from 0.933 to 0.996, while the face-only approach yielded slightly lower accuracies (0.868 to 0.923). For symptom classification, the full-body approach achieved accuracies between 0.607 and 0.650, outperforming previous work using radiomic features on the same dataset. The face-only approach for symptom classification resulted in accuracies ranging from 0.544 to 0.612, still demonstrating improvement over prior results. The study concludes that the proposed methodology is effective for classification tasks on thermal images, particularly for gender classification. While the results for symptom classification are not yet reliable enough for standalone diagnostic use, the high sensitivity scores suggest potential as a screening tool. The research highlights the promise of infrared thermography combined with machine learning for medical applications, especially in scenarios where traditional imaging methods are impractical or pose risks due to radiation exposure. Future work could explore data augmentation, additional patient data, and applications in other medical domains.