The LIRIS laboratory leads researches on the analysis of shapes
extracted from images. In particular in the ANR ReVes project, which goal
is to develop a recognition application of plants on Smartphone, studies
deals with general leaf shapes and characterization of apical, basal and
margin areas to recognize tree species. Therefore, the LIRIS laboratory
studies image analysis methods that could be applied, among other
things, to the recognition of plants. This can occur in natural environment,
such as forests or grasslands (in which case the species is the local flora)
but also in urban / peri-urban areas, such as gardens or public parks (in
this case, it can be imported species). This topic is one of the themes
addressed in the LabEx IMU.
Planar shapes are the subjects of many studies, based on concepts
of pattern recognition (classification, data analysis, statistics…) and
geometry (discrete, continuous, algorithm, etc.). On the one hand, this
research focuses on the representation of shapes, where one wants to
obtain the most concise and representative description. Examples of basic
descriptors include area, compactness, mean curvature (due to their
simplicity and generality, these descriptors might not be relevant in
practical cases. Often, more complex features are considered). On the
other hand, current research on planar shapes also deals with their
classification and their statistical analysis. These objectives involve a
comparison; more precisely a quantification of similarity between two
shapes. Then, it is necessary to have measures of similarity, which most
often will respect the property of invariance under certain transformations.
For example, rotation invariance is generally desired. Thus, measures that
give maximum similarity between a shape and the same one rotated by
any angle are appropriated. In addition to measuring the similarity, it is
also useful to combine two shapes or more, using interpolation methods.
For example, extraction of statistical measures on a set of similar but not
identical shapes may involve the computation of an average shape, being
considered as the representative of the set of shapes.
The interpolation of two shapes may be carried out either in the
space of features, discussed in the previous paragraph, either directly on
the shape itself. In this thesis, we propose to work in the space of shapes
and propose methods to compute representative shape of a set of shapes.
To this purpose, several representations of shapes can be used. Using
contour-based representation, algorithms work on structures representing
the linking between boundary points of the object. The contour can be
represented using a polygon (a list of real coordinates vertices) or by
means of a discrete curve (a sequence of adjacent pixels with integer
coordinates). In a region-based representation, algorithms manipulate
discrete sets (usually connected) of pixels. In this thesis, we will explore
methods of combination that can use contour and/or region aspects.
Often, planar shapes are the result of a segmentation step, which
generates possibly noisy or incomplete shapes. Indeed, various
phenomena can affect the quality of the image: noise, blur, partial
occlusion, etc. So we will focus, in a first time, on segmentation methods
that can integrate knowledge in order to reduce errors caused by these
phenomena. In a second step, we will study methods of combinations of
shape, taking into account noise or missing parts.