Machine Learning

Research at CLIR in machine learning covers a wide range of topics going from applications of classical machine learning algorithms for solving challenging problems in fields such as computer vision, robotics and pattern analysis of bio-metric data, to the theoretical study and design of new learning models.


 

Chagas Parasite Segmentation

Chagas disease is a tropical parasitic disease caused by the protist Trypanosoma cruzi. About 7 million to 8 million people worldwide, mostly in Latin America, are estimated to be infected. Since blood samples analisys is time consuming, this project proposes automated systems based on convolutional neural networks to detect this parasite.

 


 

Comet Assay

The comet assay (single-cell gel electrophoresis) has been a reliable method for the study of DNA damage of the cells and the evaluation of their regenerative capacity, playing an important role for the analisys of degenerative diseases (cancer, diabetes, osteoporosis, etc.). This project proposes automatic segmentation methods to extract the DNA cells from comet assay images and to study their degradation.

 


Online Speaker Diarization

Speaker diarization is the process of partitioning the speech signal in order to group speech segments corresponding to the same speaker. This research proposes an online speaker diarization system using deep learning techniques.

 


 

 

Traffic Signal Recognition

The recognition of traffic signals is one of the fundamental tasks in the advanced driving assistance systems, since most of the actions the vehicle must take to maintain a safe and convenient driving fall on them. The project implements a deep learning algorithm (YOLO) to classify and detect traffic signals from the Yucatan state.

 


 

Lunar Crater Detection

The identification of craters on the lunar surface is high scientific interest for to obtain relevant information about the periods of meteoric incidence in the solar system. In this project, we propose machine learning methods to detect craters from images of the lunar surface

 


 

Mexican Sign Language Recognition using Machine Learning

A sign language is an effective means of communication that provides the deaf a way to interact with the world around them. Implementation of machine learning techniques to develop a sign language translator may help to bridge the gap between those who can hear and those who cannot.


 

Sign Recognition for Service Robots Communication

Service robots are aimed to assist people in daily life task. In this project, we develop different body-gestures recognition techniques to interact with service robots.

 


 

Iris Recognition based on Machine Learning Techniques

Biometrics is a discipline that studies methods for verification and identification of individuals based on physical or behavioral characteristics of a person. The iris of the human eye is a structure which can be used as a reliable feature to identify individuals since its intrinsic pattern is considered unique and permanent for each individual. In this project, an iris recognition system is proposed based on speeded up robust features (SURF) matching statistics which feed a learning algorithm to automatically decide whether the two images correspond to the same iris. Experiments performed on CASIA iris image database show that our method has promising capabilities for iris recognition.

 


 

Learning problems

Historically, the problem of supervised classification has been tackled using the neural network approach or with algorithms of a statistical nature. In particular, neural networks, when correctly designed, are capable of delivering efficient solutions to the classification problem. However the analysis of the solution and its properties is very complex and this is why they are commonly referred to as "black box" solutions. In recent years, mathematical models have been proposed based on the calculus of variations for the problem of data classification. This different approach allows us to use powerful mathematical tools such as functional analysis and differential geometry for model analysis. We are interested in developing models and fast algorithms and to tackle the many challenges that appear in doing so.