Video classification of odors with Convolutional neural networks

José Maria Lopes Rafael Ledesma Frazão

Key information

Authors:

José Maria Lopes Rafael Ledesma Frazão (José Maria Lopes Rafael Ledesma Frazão)

Supervisors:

Susana Isabel Conde Jesus Palma; Maria Margarida Campos da Silveira (Maria Margarida Campos da Silveira)

Published in

February 1, 2021

Abstract

Trying to mimic natural olfactory systems, by achieving the perfect combination between chemical sensing and artificial intelligence, has been an important task among the science community. Over the last decade, e-nose technologies have undergone important developments on their way to be implemented in real life situations such as food monitoring, air quality monitoring, etc. For that to be possible, e-noses need to possess a very efficient and accurate pattern recognition component. In this work I will investigate how CNN based algorithms can be implemented in these devices to help them achieve good results. With deep learning being a hot topic nowadays, and with the advance in computational hardware, plenty of very good architectures have emerged that can achieve outstanding results in pattern recognition tasks. Different architectures are approached, from simple 3 dimension CNNs to an object detection algorithm (YOLO) and recurrent algorithms capable of handling sequencial data (LSTMs). Since this work focuses solely on the pattern recognition side of e-noses, it will work on an already produced dataset. At first, a gel was exposed to 11 different volatile organic compounds and its reaction to them was recorded in the format of image sequences. The objective is to find the algorithm capable of classifying the image sequences corresponding to the the 11 VOC classes with the highest possible score. On the second task, this gel was exposed to different concentrations of the same compound. The second goal is to find the model best capable of predicting the actual exposed concentration.