SGN-81006 SIGNAL PROCESSING INNOVATION PROJECT, 5-8 CR

2017-2018 Fall implementation on periods 1-2

Project Topics

Cancer cells image data collection, segmentation and analysis Sipilä Vili has reserved this topic, but there is still room for another student in the project.
Mammography Image Analysis Methods for Breast Cancer
Tracking Temporal Color Constancy Using Kalman Filter This topic is reserved by Xing Wenzhu.
Optimal Life with IT support Program: Topic 1. Exercising with You Tube videos and not alone
Optimal Life with IT support Program: Topic 2. Walking with companion tracking app
Optimal Life with IT support Program: Topic 3. Phone usage addiction support tool
Urban Informatics Project
Scalable representation of Light field images This topic is reserved by Bai Jingyi and Liu Yanni.
Environment classification based on GNSS and sensor characteristics Välkki Inkeri has reserved this topic, but there is still room for 1-2 students in the project.
Recognition system for animal vocalization imitation and prototype game design This topic is reserved by Kinnunen Nyyti, Saarimäki Ansse and Parviainen Eveliina.
Open topic related audio processing and analysis
Sound recognition system with Tensorflow on Android platform This topic is reserved by Pohjola Aku.
Analysis tools for bat sonar signals
Automatic Classification of Benthic Macroinvertebrates Tran Thanh Dat and Tuhola Matti have reserved this topic, but there is still room for one more student in the project.
Estimation of signal attenuation model parameters from uncertain data
Distance Measures of Probability Distributions for Target Recognition
Indoor localisation using aroma fingerprints
Biological tissue classification from ion mobility spectrum's with machine learning


Cancer cells image data collection, segmentation and analysis

Sipilä Vili has reserved this topic, but there is still room for another student in the project.

Description:
The Molecular Signaling Lab (MS Lab) combines Cell and Molecular biology, image and signal processing fields to understand signaling networks of cancer cells. The proposed project, integrating ongoing multidisciplinary work, will be composed of: 1) Wet lab experimentations including the techniques of 2D culture of cancer cells and confocal microscopy imaging, 2) microscope image analysis and data extraction. The project students will be handling 2D culture, imaging and analyse the images. For more details, please contact the client.

Requirements:
1-2 students with a interest in lab experiments and/or imaging and computational system biology. Knowledge in both fields is an advantage and some experience in image anlysis is required.

Contact:
Meenakshisundaram Kandhavelu, meenakshisundaram.kandhavelu
at tut dot fi
 
For an outline of the group’s research:
http://www.tut.fi/ms-lab/

Mammography Image Analysis Methods for Breast Cancer

Breast segmentation


Anatomical mapping

Feature extraction


Description:
Computerized algorithms for the analysis of mammography images are of key importance for the assessment and detection of breast cancer. With this project, the student will learn and develop different tools for fully-automated analysis of breast images in order to tackle different problems such as breast segmentation, cancer detection and anatomical breast mapping.

Available topics:
1. Breast segmentation: Detection of the breast contour and pectoralis line
2. Anatomical mapping: Nipple detection and geometric representation of breast anatomy
3. Feature extraction: Application of texture analysis algorithms and machine learning techniques for cancer detection and diagnosis.
This third topic is reserved by Ghazi Pedram and Monakhov Dmitrii.

Requirements:

Fundamentals of signal processing, experience in programming (Python, C++, Matlab)

Supervisor:
Said Pertuz (said.pertuz
at tut dot fi)

References:
[1] S. Pertuz, C. Julia, D. Puig, A novel mammography image representation framework with application to image registration, Proc. International Conference on Pattern Recognition ICPR 2014, pp. 3292-3297
[2] Y. Zheng et al., Parenchymal texture analysis in digital mammography: a fully automated pipeline for breast cancer risk assessment, Medical Physics 42(7):4149-4159, 2015.

Tracking Temporal Color Constancy Using Kalman Filter

This topic is reserved by Xing Wenzhu.

Background:
The human visual system perceives colors of objects independently of the incident illumination under varying conditions. This ability is known as color constancy. Why a machine (camera) cannot have this ability? This problem is a key problem in computer vision and graphical application, e.g. fine-grained classification, semantic segmentation and scene rendering. The problem of estimation illumination in a single image has faced difficult challenges, but that of estimation in videos (temporal sequences) has received minimum attention.  This is the reason we choose temporal color constancy as our target.

The projects aims to study how a Kalman filter algorithm do illumination estimation based on a series of frames. In others words, after studying a series of consecutive frames, the Kalman filter learns the trajectory of illumination change based on image content change. The Kalman filter, known for producing accurate estimates based on a series of observations over times, not based on a single measurements alone, will be widely adopted in this project.

Objectives of the project can be tuned based on the interest of the student and the discussion of the both sides. For example, the student comes out of tracking method more suitable to the color constancy problem.

Objectives:
1) Basic literature survey of Kalman Filter and its family members, e.g. Particle Filter, Extended Kalman Filter and Unscented Kalman Filter.
2) Success Validation of Kalman Filter on a simple example dataset. This dataset will be provided by the client.
3) Find suitable apperance features which suits color constancy, e.g. color histogram in log space.
4) Compare Kalman Filter’s performance with other temporal methods on some public convincing dataset, e.g. SFU grey ball dataset.

Group size:
1 person

Required deliverables:
1) Code (Python)
2) Report including documentation the used methods and the results of the study
3) If this project goes smoothly and the results are convincing and novel, then we head to a submission to International Conference on Pattern Recognition (ICPR 2018) held in Beijing.

Prerequisites:
1) Basic C++ skills)
2) Basic Python skills

Clients:
Yanlin Qian and Prof. Joni Kamarainen
Labrotory of Signal Processing, TUT
firstname.lastname at tut dot fi
Rooms TF316

References:
Kalman filter, Wiki Page
Wan E A, Van Der Merwe R. The unscented Kalman filter for nonlinear estimation[C]//Adaptive Systems for Signal Processing, Communications, and Control Symposium 2000. AS-SPCC. The IEEE 2000. Ieee, 2000: 153-158.
Dongliang Cheng, et.al.,  Effective Learning-based Illumination Estimation Using Simple Features, CVPR 2015
Jonathan Barron, et.al., Convolutional Color Constancy, ICCV 2015.
Jonathan Barron, et.al., Fast Fourier Color Constancy, CVPR 2017.
Yanlin Qian, et.al., Recurrent Color Constancy, ICCV 2017.
You find it.

Optimal Life with IT support Program: Topic 1. Exercising with You Tube videos and not alone

This program is for optimizing various life issues using IT support. Among those issues are lack of physical effort and addiction to IT devices. IT devices may help and there are applications which offer support but most people are not able to use for longer time. We identified reason for this is loneliness and lack of social interaction and we intend to develop apps dealing with this problem.

Description of Topic 1:
Lack of physical effort and insufficient energy expenditure comparing to the calories consumption are major issue in developed countries. Remedy for this is regular exercising for maintaining fitness but that is not easy since it requires strong motivation. There are now many devices and applications intended for guiding exercising but practice shows that in general people do not use them for longer time since they lack support. This project aims for providing support for the case of fitness videos available on You Tube (look e.g. Fitness Blender there). Suppose that somebody wants to make a regular routine using You Tube exercise videos. Doing this alone is hard since it wears off motivation. To prevent this, one has to have companion(s) with which information about the execution of particular video exercises will be exchanged. That is, after running a video exercise from You Tube a message will be send automatically to the companion. In this way, people involved in exercising are motivating each other to keep with the routine. They can synchronize their efforts by executing the same video exercise while not being in the same place. The project requires thus implementation of an overlay application for You Tube videos which will register which exercise was executed and send a message to the companion application about this. Demo implementation can be done in a browser and/or Windows and/or Android.

Requirements:
1-2 persons depending on the scope of demo implementation. Knowledge in the Web and/or Windows and/or Android programming is needed.

Contact:
Irek Defée, irek.defee (at) protonmail (dot) com, sms 0400736612

Optimal Life with IT support Program: Topic 2. Walking with companion tracking app

This program is for optimizing various life issues using IT support. Among those issues are lack of physical effort and addiction to IT devices. IT devices may help and there are applications which offer support but most people are not able to use for longer time. We identified reason for this is loneliness and lack of social interaction and we intend to develop apps dealing with this problem.

Description of Topic 2:
There are many applications which intend to overcome sedentary lifestyle by encouraging people just for little walking outside each day. These applications use GPS and/or movement sensor in mobile devices to track walking and accept walking period after detecting e.g. 10 minutes of brisk movement.

Ideally it would be best if an app of this kind stimulates people to regular waling over longer period of time, in the end a new habit of walking would be created. It turns out however that most people use such apps at most for a couple of weeks. One reason for this is that even with the app people are left alone while they need to interact with other persons in this.

Here we try to solve this problem by creating an app which will communicate with another person for exchanging data about walking. The data are compared and the persons can support and stimulate each other to maintain regular schedule.

The app will be implemented in Android starting from open source walking demo app. Data exchange communication system and data comparison system is going to be designed and implemented.

Requirements:
1-2 persons depending on the scope of demo implementation. Knowledge of Android programming is a plus but learning it during the project is possible.

Contact:
Irek Defée, irek.defee (at) protonmail (dot) com, sms 0400736612


Optimal Life with IT support Program: Topic 3. Phone usage addiction support tool

This program is for optimizing various life issues using IT support. Among those issues are lack of physical effort and addiction to IT devices. IT devices may help and there are applications which offer support but most people are not able to use for longer time. We identified reason for this is loneliness and lack of social interaction and we intend to develop apps dealing with this problem.

Description of Topic 3:
Many people are using their phones obsessively, wasting a lot of time. In the end it becomes an addiction which they cannot overcome even if they decide to deal with it. There are many applications trying to solve this by monitoring and reporting how much user is using the phone and some applications allow to put a limit, locking the phone after the limit is used. But, like with many other addictions, dealing with it alone is very often bound to be failure. Modern approaches to addiction put emphasis on support involving groups of persons having the same problem. Hence, here we would like to develop application monitoring the usage of the phone and allowing exchange of usage data with other people trying to solve this problem. Then, people can offer mutual support to keep with promises of reducing the usage. There are here obvious issues with data privacy but they can be solved with anonymization.

Requirements:
1-2 persons depending on the scope of demo implementation. Knowledge of Android programming is a plus but learning it during the project is possible.

Contact:
Irek Defée, irek.defee (at) protonmail (dot) com, sms 0400736612

Urban Informatics Project

It is possible to pass this course in the form of Urban Informatics project that is done together with architecture students.

Urban informatics course will be third time organized together with the Architecture Lab of TUT (Prof. Harry Edelman). The goal is to find a problem where you can provide architecture students data analysis for their needs and real data will be collected from the city centre. You will receive 5-8 cr plus in the best case one publication about this new topic of "urban informatics".

Check also this course "ARK-53806 Sustainable Design Studio" and enrol to its Moodle to find details of the project.

Go to the first lecture of the course
ARK-53806, next week Tuesday 29th at 12am in theatre RJ108 in the Architecture Building.
 
For more information:
Joni Kämäräinen
joni.kamarainen at tut dot fi

Scalable representation of Light field images

This topic is reserved by Bai Jingyi and Liu Yanni.

Background:
Light Field capture systems can be created from an array of cameras, each seeing the scene from a slightly different perspective and contributing a varied vantage point. Every vantage point in the array is then combined to produce Light Field images. That is, the light field is a vector function that describes the amount of light flowing in every direction through every point in space. With the plenty of information, we can generate images of different focal points, variable apertures, shift of perspective and even simulate focal plane effects. (Examples available on: http://blog.lytro.com/what-is-light-field/) However, the plenty of information also introduces a huge amount of data for storage. Thus, an efficient compression method is required, which is also the target of this project.

Objectives:
1)    Literature review of Light field image representation and compression.
2)    Implementation of benchmark method, i.e. published work related to the topic.
3)    Explore novel representation method for Light field images, with the guidance of the client.
4)    Implement the novel method and compare its performance with benchmark.

Group size:
1-2 person

Prerequisites:
1)    Experience in programming (C++, Matlab).
2)    Preferable to have some knowledge in deep learning, but not necessary.

Client:
Li Yu (li.yu at tut dot fi)
Room TF409

Environment classification based on GNSS and sensor characteristics

Välkki Inkeri has reserved this topic, but there is still room for 1-2 students in the project.

Description:
The target accuracy of a GNSS solution often correlates with the environment which the GNSS receiver is expected to operating within. In real-world scenarios the observed GNSS characteristics of the environment often change rapidly resulting in the expected performance of the receiver to change as well. It is the goal of this project to determine a methodology for characterizing environments based on their GNSS (and optionally sensor) characteristics such that accuracy targets can be aligned with the observed performance of the receiver.

How will the output of this project be used in practice?
u-blox has accuracy targets for various receiver configurations used in a given environment (E.g.,  less than 1m accuracy for a high precision receiver used on open highway). As u-blox is interested in how our receivers perform in all environments we collect data with a test vehicle across a wide range of environments (e.g., highway, tunnels, rural, urban, deep urban, ...) and often a single log file contains multiple environments. In order for u-blox to evaluate if our receivers perform as expected we must divide the recorded data to meaningful segments such that the performance can be compared with a relevant target for each segment. Currently the method for segmenting the data is a time-consuming, manual process involving the plotting of the driven route on Google Earth and recording the time stamp of roughly when the environment changes from one category to another.

The challenges we are trying to overcome are the following:
  1. Lack of well defined environment categories. Without well defined categories individuals can have different interpretation of the environment based on their own experiences. Today we use rural, urban, highway, tunnel and other. The goal is to have a (new) set of environment categories that can be described by their observed GNSS and sensor characteristics. The categories should also relate to visually observed characteristics that could be understood by our customers.
  2. Subjective decision on when environment changes from one to another. Our current approach is based on visual inspection of a route on Google Earth that may not be up-to-date. Having a solution based on observed GNSS and sensor characteristics would ensure consistency on how recorded routes are divided by category across all collected data by different people.
  3. In-efficient use of time categorizing data: The process of splitting recorded logs to various environments is time consuming especially for hundreds of hours of recorded data:
    1. Convert receiver logs to kml format for plotting on Google Earth
    2. Open and plot kml file on Google Earth
    3. visually scan route on map looking for rough location when environment changes based on loose definition
    4. record time stamp of environment change
Requested output of project:
u-blox's contribution:

Group Size:
1-3 students - number of students would impact how complex of a solution could be developed and the amount of data that could be analyzed in order to verify the results.

Client:
u-blox Espoo Oy; Tampere office
Michael Schmidt
mike.schmidt at u-blox dot com

Recognition system for animal vocalization imitation and prototype game design

This topic is reserved by Kinnunen Nyyti, Saarimäki Ansse and Parviainen Eveliina.

Background:

Humans have common onomatopoeias for animal vocalizations. These onomatopoeias are to imitate animal sound, and they often language dependent. For examples, dog barking in Finnish is commonly imitated with "hau-hau", whereas in English "woof-woof" is used and in Mandarin "wang-wang". These onomatopoeias are used in children games when teaching animals species and referring to them.

Objectives:
1) Collect small sound dataset with limited set of animal onomatopoeias. For example, 2-8 people imitating dog, cat, cow, and sheep vocalizations a few times is sufficient.
2) Based on the collected dataset, a simple sound recognizer system is developed to recognize these sounds. Client will provide development framework for sound recognizer development. The recognizer system should reach acceptable performance level to allow game prototyping in later stages.
3) Develop a simple game idea around animal onomatopoeia recognition.
4) Implement prototype game with the sound recognizer system. The game should capture audio real-time.

Group size:
1-3 persons

Depending on the group size, or student's interest, project can be steered more towards sound recognition system development or towards game design. The project work can be easily split between students so that one is concentrating on sound recognition system and one on game design and implementation.

Required deliverables:
1) Sound dataset animal onomatopoeias
2) Sound recognizer system (Python code)
3) Prototype game build around sound recognizer (Python code)
4) Project documentation

Prerequisites:
1) Basic signal processing and machine learning skills
2) Python skills

Clients:
Toni Heittola and Tuomas Virtanen
Department of Signal Processing, TUT
firstname.lastname at tut dot fi
Rooms TC338 and TF311

References:
Sound recognition framework: https://github.com/TUT-ARG/DCASE2017-baseline-system
Game framework: https://www.pygame.org

More information:
Toni Heittola
toni.heittola at tut dot fi

Open topic related audio processing and analysis

Background:
This is an open topic related to audio signal processing and/or content analysis of sounds. The topic will be selected together with student, and can be customized based on student's interest quite widely. The topic can be selected, for example, based on interesting application utilizing audio, or based on interesting new approach on audio processing or sound analysis.

Audio signal processing techniques can be found in many applications. Audio processing techniques include, for example, voice conversion (converting male speech to female), sound synthesis, audio effects, echo cancellation, sound source separation (separating specific sound from a background).

The content analysis of sounds uses usually techniques based on audio classification, sound source separation, automatic music transcription, self localization, speaker tracking, or speech recognition. Applications include context aware devices, media retrieval from large datasets (content search), human-machine interaction, analysis of social situations, life-logging, to name a few.

Objectives:
1) Select topic and conduct brief literature review on state-of-the art technical approaches
2) Select suitable state-of-the-art approach for more detailed study
3) Implement prototype system with selected approach
4) Evaluate the performance of the approach
5) Report findings

Group size:
1-2 persons

Required deliverables:
1) Brief literature review
2) Implementation source code (preferably released as open source code)
3) Project documentation

Prerequisites:
1) Basic signal processing and/or machine learning skills
2) Basic Python or Matlab skills

Clients:
Tuomas Virtanen and Toni Heittola
Department of Signal Processing, TUT
firstname.lastname at tut dot fi
Rooms TF311 and TC338

More information:
Toni Heittola
toni.heittola at tut dot fi

Sound recognition system with Tensorflow on Android platform

This topic is reserved by Pohjola Aku.

Background:
In recent years, TensorFlow by Google has gained popularity for both machine learning research and deployment of machine learning applications to the market. TensorFlow can be used on Android platform, however, most of the applications using it are concentrated on image recognition. Recently, Google has released TensorFlow Lite to Android platform to enable the development of neural networks optimized for computational resources available on mobile phone.

In this project, the student will investigate TensorFlow usage on Android platform by building simple real-time or near real-time sound recognition application. Sound to be recognized can be either acoustic scenes (street, office, etc.) or single sound class such as dog barking or baby cry.  The neural network used in the recognizer will be trained outside the phone on Python platform and moved to the recognizer. Client will provide Python framework to train neural networks, and pre-trained models (Keras models which are wrapping TensorFlow model) to bootstrap the development.

If time permits, the student should investigate both TensorFlow and TensorFlow Lite usability in the project.

Objectives:
1) Build prototype application for sound recognition on Android platform with minimal GUI
2) Develop work flow to ensure acoustic feature extraction is working similarly on offline training (Python) and on online recognition (Java)
3) Study TensorFlow Lite, identify the differences to TensorFlow and see how can be utilized in prototype application.
4) Verify that system recognizes sounds
5) Report findings

Group size:
1 person

Required deliverables:
1) Source code
2) Prototype application
3) Project documentation

Prerequisites:
1) Basic signal processing and machine learning skills
2) Familiar with Android development environment
3) Java skills
4) Basic Python skills

Clients:
Toni Heittola and Tuomas Virtanen
Department of Signal Processing, TUT
firstname.lastname at tut dot fi
Rooms TC338 and TF311

References:
https://www.tensorflow.org/
https://github.com/amir-abdi/keras_to_tensorflow
https://blog.mindorks.com/android-tensorflow-machine-learning-example-ff0e9b2654cc

More information:
Toni Heittola
toni.heittola at tut dot fi


Analysis tools for bat sonar signals

Description:
The task is to develop MATLAB-tools for segmentation and TF (time frequency) analysis of sonar signals of some bat species. The dataset contains about 150 recordings and consists of echo location sounds of probably 5 different species of bats.
The tools needed:
- An automatic segmentation tool
- MATLAB functions that compute besides spectrogram also Wigner-Ville and pseudo Wigner-Ville transformations of the signals
Some preliminary work has been done using the TF-toolbox by Flandrin et al. (1996a,b) but it is not completely compatible with current MATLAB versions.  

Our research question is whether these 5 species of bats can be recognized by machine using only the sounds. The students are welcome to present also completely different solutions to this research question.  

Group size : 2-3 students

Clients name: Juha T. Tanttu (juha.tanttu
at tut dot fi)
The client works at TUT Pori campus. So the client is available mostly via email and other means of remote access.

References:
Flandrin, P., & Lemoine, O. (1996). Time-Frequency Toolbox, 1995–1996.
François Auger, Flandrin, P., Gonçalvès, P., & Lemoine, O. (1996). Time-Frequency Toolbox for Use with Matlab - Reference Guide. October, 1995–1996. Retrieved from http://tftb.nongnu.org/
Flandrin, P. (1998). Time-frequency/time-scale analysis (Vol. 10). Academic press.

Automatic Classification of Benthic Macroinvertebrates

Tran Thanh Dat and Tuhola Matti have reserved this topic, but there is still room for one more student in the project.

Description:

Managing the water quality of freshwaters is a crucial task worldwide. One of the most used methods to biomonitor water quality is to sample benthic macroinvertebrate communities, in particular to examine the presence and proportion of certain species. Currently, the benthic macroinvertebrate samples are labeled manually by human experts. Automating the classification step could therefore safe a significant amount of manual effort.

There are some challenges in automatic classification of benthic macroinvertebrates.  Different species are hard to discriminate and thus the classification task is really fine-grained. Typically, the cameras and imaging setup used for new samples of macroinvertebrates are not identical. A classifier trained using images taken in different circumstances may fail. Furthermore, the classes are highly-imbalanced, while the most interesting species are the ones seldom found. In this project, solutions to these problems are considered. The specific project objectives will be agreed with the student(s).

Group size:
1-3 persons

Prerequisites:
1) Programming skills
2) Knowledge of classification techniques
3) Preferably earlier experience of applying some public deep learning tools

Client:
Jenni Raitoharju
Department of Signal Processing, TUT
firstname.lastname at tut dot fi
Room TF404

Estimation of signal attenuation model parameters from uncertain data

Description:
Path loss models are used to model how a signal attenuates in the path from the transmitter to receiver. Using RSS (Received Signal Strength) values from multiple transmitters and corresponding path loss models one can estimate a user location. RSS based localization can be used indoors where GPS is not available. The parameters of a path loss model vary for different transmitters and have to be estimated. The estimation of path loss parameters can be done based on multiple RSS measurements at different locations. However, in indoor crowdsourcing the locations usually cannot be known exactly nor with errors until the path loss models have been constructed. To overcome this problem the user locations can be estimated based on pedestrian dead reckoning. Pedestrian dead reckoning uses inertial sensors that provide relative motion. The initial state for pedestrian dead reckoning can be obtained using GPS measurements outdoors. However, when estimating the user locations using the pedestrian dead reckoning the location errors accumulate. In this project work, the goal is to estimate the path loss model parameters from data, which contains locations that have accumulated errors.
 
Group size: 1-2 persons
 
Prerequisites:
1) Programming skills
2) Knowledge of multivariate statistics mean and covariance
3) Basic knowledge of optimization methods
 
Client:
Matti Raitoharju

firstname.lastname at tut dot fi

Distance Measures of Probability Distributions for Target Recognition

Background:
Many practical applications rely on target recognition technology, where computers automatically detect and recognize targets based on sensor measurements. In model-based target recognition approaches, the sensor response of the potential targets is simulated based on their physical models. The simulated response of the modeled targets is then used as the hypothesis for different targets in the interpretation of measured response. In some applications, the target signature is greatly influenced by the target aspect angle, and consequently the uncertainty with the estimation of the target aspect needs to be considered by the recognition algorithm.

Objectives:
1) Implement one or more statistical distance measures to compare probability distributions. The distance measurements aim at quantifying the dissimilarity between the measured and modeled target signatures in various aspect angles. Measurements of the target are accumulated over time, so the distance between the measured and modeled target distribution evolves as new observations are obtained.
2) Develop and implement one or more simple target recognition algorithms based on the calculated distance measures.
The client will provide the sensor measurements as well as the modeled sensor responses of the reference targets.

Group size:
2–4 persons
The extent of the project work (e.g. the number of distance measures and recognition algorithms compared) will be scaled depending on the group size.

Required deliverables:
1) Target recognition system (MATLAB code)
2) Project documentation

Prerequisites:
1) Basic machine learning skills
2) MATLAB skills
3) Fluent skills in both written and spoken Finnish

Clients:
Minna Väilä and Marja Ruotsalainen
Laboratory of Signal Processing, TUT

References:
Statistical distance: https://en.wikipedia.org/wiki/Statistical_distance

More information:
Minna Väilä
minna.vaila at tut dot fi

Indoor localisation using aroma fingerprints

Description: Indoor localisation nowadays uses various networks and sensors. One sensor type that has not been studied yet thoroughly for the purpose of localisation are ion mobility spectrometry (IMS) sensors. IMS-based devices have been successfully used for detecting and classifying, for example, hazardous objects. The purpose of this project is to study if it is possible to distinguish different locations inside a building based only on their smell. We will provide the students some earlier IMS measurements, but the students are also expected to collect new measurements themselves. After the data collection the students should test several classifiers for localisation (Matlab toolbox can be used). This means, the data should be divided into a training set (used for training the classifiers) and a test set (used for testing the classifiers). The students are expected to write a report about their findings, which explains where and under which circumstances the data has been collected, which classifiers have been tested and how they work, and a thorough discussion of the results.

Group size: 1-3 students (teams of 2-3 preferred)

Requirements: Students should have a background in data engineering and machine learning. The focus of the project is on supervised learning.

Contact: philipp.muller(at)tut.fi

Biological tissue classification from ion mobility spectrum's with machine learning

Odors can be used for determining the patient’s state of health. We measure odors with differential mobility spectrometer (DMS). The spectrum, more precisely a dispersion plot, is similar to an image; therefore image-based classification can be effective in distinguishing different tissue types from each other.

Students should study what are the optimum parameters for existing classification methods: (KNN, LDA), and possibly test new methods.
(Support vector machines, naïve Bayesian etc.)

For more credits there is a possibility to include more complex properties, such as feature extraction and addition of parallel information for classification.

Group size: 1-3 students (teams of 2-3 preferred)

Requirements: Students should have a background in data engineering and machine learning. The focus of the project is on supervised and semi-supervised learning.

Contact: anton.kontunen
(at)tut.fi


Main page