Human odour detection approach using machine learning
Recognizing the human considered as old and contemporary task. This problem is now solved by using biometrics. Technically biometrics is " the automated technique for measuring an individual's physical or personal characteristic and comparing it to a comprehensive database for identificati...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/35177/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/35177/2/FULLTEXT.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Recognizing the human considered as old and contemporary task. This problem is now solved by using biometrics. Technically biometrics is " the automated technique for measuring an individual's physical or personal characteristic and comparing it to a comprehensive database for identification purposes". This thesis presents a problem with the selection of appropriate human (Volatile Organic Compounds) voes emitted from sweat for human odour classification, all gasses emitted by humans through sweat have been collected and detected using the latest technology (High Resolution GCMS / TOF) Gas Chromatograph Mass Spectrometry/Time of Flight. Different people (15 people) with different ages and genders have been tested, some people have been tested several times. There is a total of 198 voes detected and methods for selecting features are used to determine which VOCs are suitable for classifying human odour. Two feature selection methods Entropy and Chi Square tests were used to identify and determine the best and most acceptable voes. There is a total of 16 stable voes extracted from 198 voes on the basis of the results obtained. In addition, 10 gasses are detected with zero values for both the entropy and the chi- square test, and these gasses are the strongest candidates to detect and classify odours. The results of this work can be used to classify specific voes for the detection of humans by odour. In this thesis, a framework for gender recognition is proposed based on human odour. 20 samples of human odour from male and female are collected, several different activation functions of the neural network (e.g., backpropagation of Levenberg-Marquardt, backpropagation of gradient descent and resilient backpropagation) and several different topologies of the neural network are tested. It is also found that with 2 hidden layers with more neurons in the hidden layers (16 and 16 neurons in which the hidden layer is) Levenberg-Marquardt was able to achieve a higher performance accuracy of 100%. The main investigations conducted in this thesis which is Human Identification from body odour followed by an investigation to prove stability and rigidity of person identification main findings. A framework for human identification is proposed distinctively based on specific human odour features. 15 samples of female and male human odour are collected from different age groups, severa I diverse functions of neural network activation are tested such as Gradient descent backpropagation, Levenberg-Marquardt back propagation, and Resilient backpropagation. Besides, numerous neural network topologies are tested by means of a selection of number of neurons and hidden layers. Different activation functions were tested TANSigmoid transfer, Linear transfer, and LOG-Sigmoid transfer. Considering the obtained results, employing two hidden layers with more neurons in the hidden layers- to be specific: 15 neurons in every layer- has yielded better accuracy in performance with an accuracy rate of 100%. The unsurpassed framework for learning algorithm to be used for human identification is Levenberg-Marquardt backpropagation learning algorithm. The best function for activation established in this research is the function of TAN- Sigmoid transfer. Finally, we investigate the effects of missing gases in human odour sample to evaluate the accuracy of classifying individual person. These missing values will be replaced by Random number between O and 1 as our research prove, the best accuracy result when missing values are introduced in the odour dataset is the Ensemble Bagged Trees. |
|---|