Colour consistency in computer vision : a multiple image dynamic exposure colour classification system : a thesis presented to the Institute of Natural and Mathematical Sciences in fulfilment of the requirements for the degree of Master of Science in Computer Science at Massey University, Albany, Auckland, New Zealand
Loading...
Date
2016
DOI
Open Access Location
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Massey University
Rights
The Author
Abstract
Colour classification vision systems face difficulty when a scene contains both very
bright and dark regions. An indistinguishable colour at one exposure may be
distinguishable at another. The use of multiple cameras with varying levels of
sensitivity is explored in this thesis, aiding the classification of colours in scenes with
high illumination ranges. Titled the Multiple Image Dynamic Exposure Colour
Classification (MIDECC) System, pie-slice classifiers are optimised for normalised
red/green and cyan/magenta colour spaces. The MIDECC system finds a limited section
of hyperspace for each classifier, resulting in a process which requires minimal manual
input with the ability to filter background samples without specialised training. In
experimental implementation, automatic multiple-camera exposure, data sampling,
training and colour space evaluation to recognise 8 target colours across 14 different
lighting scenarios is processed in approximately 30 seconds. The system provides
computationally effective training and classification, outputting an overall true positive
score of 92.4% with an illumination range between bright and dim regions of 880 lux.
False positive classifications are minimised to 4.24%, assisted by heuristic background
filtering. The limited search space classifiers and layout of the colour spaces ensures the
MIDECC system is less likely to classify dissimilar colours, requiring a certain
‘confidence’ level before a match is outputted. Unfortunately the system struggles to
classify colours under extremely bright illumination due to the simplistic classification
building technique. Results are compared to the common machine learning algorithms
Naïve Bayes, Neural Networks, Random Tree and C4.5 Tree Classifiers. These
algorithms return greater than 98.5% true positives and less than 1.53% false positives,
with Random Tree and Naïve Bayes providing the best and worst comparable
algorithms, respectively. Although resulting in a lower classification rate, the MIDECC
system trains with minimal user input, ignores background and untrained samples when
classifying and trains faster than most of the studied machine learning algorithms.Colour classification vision systems face difficulty when a scene contains both very
bright and dark regions. An indistinguishable colour at one exposure may be
distinguishable at another. The use of multiple cameras with varying levels of
sensitivity is explored in this thesis, aiding the classification of colours in scenes with
high illumination ranges. Titled the Multiple Image Dynamic Exposure Colour
Classification (MIDECC) System, pie-slice classifiers are optimised for normalised
red/green and cyan/magenta colour spaces. The MIDECC system finds a limited section
of hyperspace for each classifier, resulting in a process which requires minimal manual
input with the ability to filter background samples without specialised training. In
experimental implementation, automatic multiple-camera exposure, data sampling,
training and colour space evaluation to recognise 8 target colours across 14 different
lighting scenarios is processed in approximately 30 seconds. The system provides
computationally effective training and classification, outputting an overall true positive
score of 92.4% with an illumination range between bright and dim regions of 880 lux.
False positive classifications are minimised to 4.24%, assisted by heuristic background
filtering. The limited search space classifiers and layout of the colour spaces ensures the
MIDECC system is less likely to classify dissimilar colours, requiring a certain
‘confidence’ level before a match is outputted. Unfortunately the system struggles to
classify colours under extremely bright illumination due to the simplistic classification
building technique. Results are compared to the common machine learning algorithms
Naïve Bayes, Neural Networks, Random Tree and C4.5 Tree Classifiers. These
algorithms return greater than 98.5% true positives and less than 1.53% false positives,
with Random Tree and Naïve Bayes providing the best and worst comparable
algorithms, respectively. Although resulting in a lower classification rate, the MIDECC
system trains with minimal user input, ignores background and untrained samples when
classifying and trains faster than most of the studied machine learning algorithms.
Description
Keywords
Computer vision, Colour vision, Image processing, Digital techniques, Research Subject Categories::TECHNOLOGY::Information technology::Image analysis