Machine Learning Model for Prediction of Human Skin Locations and Conditions through Emissivity Measurements رسالة ماجستير

Abstract

In this thesis, machine-learning techniques were used to predict the human skin emissivity and to classify the human skin status based on its emissivity for wet and dry skin conditions. Predicting skin emissivity was made by selecting a specific skin location and using other locations from the same person to predict the selected location emissivity measure. Predicting measurement location emissivity will help to determine the normal emissivity value for the needed location; the implication of having this is the non-invasive diagnosis of diseased skin as the location of the skin may be infected with one of the skin diseases or thermal burns or may be covered with a hidden object, which will affect the emissivity. Regression models were used to predict the emissivity of the skin from nine locations using Linear Regression, K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Multiple-Layer Perceptron Neural Network (MLPNN). Experimental results show that MLPNN gives the highest accuracy for most measurement locations, and the palm of hand achieved the highest accuracy of 91.3%. and the lowest mean absolute error of 0.0184 for the elbow location using MLPNN. For the second part of the development which is the classification of the skin conditions; many algorithms were used to predict the skin status based on the emissivity of the dry and wet skin conditions, and those are KNN, random forest, decision tree, and MLPNN. vi The dataset of regression was collected from previously published studies of emissivity this dataset contains 540 records for 60 volunteers with 9 different measure locations from both genders. This dataset contains 5 features, location, ethnicity, age, mean emissivity, and gender The dataset for the classification was collected using the calibrated radiometer for a sample of 120 participants from 9 different locations in dry (normal skin) and wet skin (skin after the application of water) skin conditions from the previous studies also, This data set feature contains location, ethnicity, age, skin status, emissivity, and gender. Results show that MLPNNs achieve accuracy of about 91.6%, recall 93.3%, precision 87.5%, and F-measure 90.3%, other detailed results will be discussed in chapter five.