BLUSSEAU_Samy.JPG
ABSTRACTThe present dissertation compares the human visual perception to computer vision algorithms based on a mathematical model called a-contrario theory.
To this aim, it focuses on two visual tasks that are at the same time easy to model and convenient to test in psychophysical experiments. Both tasks consist in the perceptual grouping of oriented elements, namely Gabor patches. The first one is the detection of alignments and the second one extends to curves, that is to say to more general arrangements of elements in good continuation. In both cases, alignments and curves, psychophysical experiments were set up to collect data on the human visual perception in a masking context.
The non-accidentalness principle states that spatial relations are perceptually relevant when their accidental occurrence is unlikely. The a-contrario theory is a formalization of this principle, and is used in computer vision to set detection thresholds accordingly. In this thesis, the a-contrario framework is applied in two practical algorithms designed to detect non-accidental alignments and curves respectively. These algorithms play the part of artificial subjects for our experiments. The experimental data of human subjects is then compared to the detection algorithms on the very same tasks, yielding two main results.
First, this procedure shows that the Number of False Alarms (NFA), which is the scalar measure of non-accidentalness in the a-contrario theory, strongly correlates with the detection rates achieved by human subjects on a large variety of stimuli. Secondly, the algorithms' responses match very well the average behavior of human observers.
The contribution of this thesis is therefore two-sided. On the one hand, it provides a rigorous validation of the a-contrario theory's relevance to estimate visual thresholds and implement visual tasks in computer vision. On the other hand, it reinforces the importance of the non-accidentalness principle in human vision.
Aiming at reproducible research, all the methods are submitted to IPOL journal, including detailed descriptions of the algorithms, commented reference source codes, and online demonstrations for each one.