Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Computer vision using biomimicking AI system

Högbom Aronsson, Elias LU (2022) EITM01 20212
Department of Electrical and Information Technology
Abstract
The aim of this thesis was to develop a biologically inspired model for unsupervised learning on visual data. A preprocessor inspired by the retina was developed, which generated event-camera like data. This can be used as a low-cost alternative to an event-camera. Further, a neuron model was developed which showed ability to learn useful features from video data in combination with the preprocessor. A number of aspects of the system were also explored to probe its properties.

A data set consisting of videos of moving 3D-printed objects was created. This was used to compare the biological model in combination with a small LSTM network against a much larger convolutional LSTM, in classifying the different objects. This gave promising... (More)
The aim of this thesis was to develop a biologically inspired model for unsupervised learning on visual data. A preprocessor inspired by the retina was developed, which generated event-camera like data. This can be used as a low-cost alternative to an event-camera. Further, a neuron model was developed which showed ability to learn useful features from video data in combination with the preprocessor. A number of aspects of the system were also explored to probe its properties.

A data set consisting of videos of moving 3D-printed objects was created. This was used to compare the biological model in combination with a small LSTM network against a much larger convolutional LSTM, in classifying the different objects. This gave promising results, where the biological model performed as well or better in most aspects, compared with the large convolutional LSTM. (Less)
Popular Abstract
To recognize things in the world is something we do effortlessly. We can recognize a bird even if we have never seen that particular species or from that angle. For computers, this is not that easy. To define everything that makes a bird, a bird, in pixel values, is not a trivial task. Modern machine learning methods try to, instead get the computer to learn all these different aspects by itself. This has proven successful for many tasks, but requires large amounts of training data and computational resources. These methods also have problems generalizing to new situations.

The company IntuiCells AI technology was built on more general assumptions and inspiration from neuroscience, to address such problems. This thesis builds on... (More)
To recognize things in the world is something we do effortlessly. We can recognize a bird even if we have never seen that particular species or from that angle. For computers, this is not that easy. To define everything that makes a bird, a bird, in pixel values, is not a trivial task. Modern machine learning methods try to, instead get the computer to learn all these different aspects by itself. This has proven successful for many tasks, but requires large amounts of training data and computational resources. These methods also have problems generalizing to new situations.

The company IntuiCells AI technology was built on more general assumptions and inspiration from neuroscience, to address such problems. This thesis builds on IntuiCell technology, to see if this novel approach works in the field of computer vision.

First, two videos were animated and rendered, to test out different aspects of the model, as wells as to prove its promise. When this was completed a recording setup, consisting of a box with controllable LEDs and a motorized slider, where objects could be moved, was constructed. From the recordings, the model was trained to recognize different objects. This gave promising results. The model could recognize the different objects, even in some new situations. This shows a first promising result, of this novel approach to computer vision. (Less)
Please use this url to cite or link to this publication:
author
Högbom Aronsson, Elias LU
supervisor
organization
course
EITM01 20212
year
type
H1 - Master's Degree (One Year)
subject
keywords
computer vision, artificial neural network, artificial intelligence, computational neuroscience, biomimicking
report number
LU/LTH-EIT 2022-864
language
English
id
9079163
date added to LUP
2022-05-12 10:50:48
date last changed
2022-05-12 10:50:48
@misc{9079163,
  abstract     = {{The aim of this thesis was to develop a biologically inspired model for unsupervised learning on visual data. A preprocessor inspired by the retina was developed, which generated event-camera like data. This can be used as a low-cost alternative to an event-camera. Further, a neuron model was developed which showed ability to learn useful features from video data in combination with the preprocessor. A number of aspects of the system were also explored to probe its properties.

A data set consisting of videos of moving 3D-printed objects was created. This was used to compare the biological model in combination with a small LSTM network against a much larger convolutional LSTM, in classifying the different objects. This gave promising results, where the biological model performed as well or better in most aspects, compared with the large convolutional LSTM.}},
  author       = {{Högbom Aronsson, Elias}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Computer vision using biomimicking AI system}},
  year         = {{2022}},
}