France: PhD Positions in Computer Science

Font size: Decrease font Enlarge font


The overall goal of the project is to improve an existing document
analysis system, which is able to convert various type of documents,
most of them originating from the French heritage and assuming various
physical forms (books,
microfilms, postcards, civil deeds, etc.). The novelty of the PhD comes
from the proposed methodology, which suggests processing images
directly in grey levels rather than in their binarized version.

Strong knowledges in image analysis and processing, as well as the
technical mastery of a programming language (such as C++ or Java) are a
must. Additional knowledges in mathematical morphology for the first
position, and experience with an environment such as Matlab or R, or
further knowledges on statistical classification/recognition methods
(SVM, neural networks, bayesian networks, …) for the second position
would be additional assets.

Interested applicants should be either French-speaking or fluent in
english. Resumes, possibly including a publication or a manuscript, and
motivation letters should be sent by e-mail only to:

l.najman[ at ] for position 1
x.hilaire[ at ] for position 2

Duration : 36 months
Salary : 39 k€ per annum.
Expected starting date : November 2008.

Position 1 : Document image enhancement using topological and morphological filters in simplicial complexes

The aim of the PhD is to improve the quality of the document images by
rectifying as much as possible the various consequences of noise, which
appear either at pixel level (white noise, locally low contrast,
blotting effect of the paper), or at character level (ink default,
paper defect, broken characters). Page level defects other than white
noise (e.g., rotation, uneven feeding of paper when using roll
scanners), however, need not be addressed.

We propose to design a morphological and topological filtering
method that operates within the framework of greylevel simplicial
complexes. Our idea is to process images at subpixel level only, and to
refrain from binarizing it. When processing printed text, the method
aims to provide an image with characters as close as possible to the
ground truth (without it be possible, however, to formally state its
performance, since the commercial OCR FineReader will be used as a
black box in the real processing chain). When coping with handwritten
text, the method shall compute, word by word, a filtered skeleton of
these words that will be used as input to feed an HCR.

In both cases, formal performances of the filtering method will be
established based on the specifications of a public OCR and HCR, and
proped up with experimental validations using FineReader and
DocumentReader in the framework of the project.

Position 2 : Contextual segmentation of document images using greylevel texture analysis
Document page segmentation is to automatically recognize and extract
its various components (text and text blocks, mathematical formulas,
halftones, captions, …).

Numerous segmentation methods are available in the literature. The
usual taxonomy grossly fits them into three families : top-down methods
(one starts from an entire page, then recursively split it until a
criterion is satisfied on each region), bottom-up methods (the opposite
approach), and hybrid methods. The latter family obviously gathers
methods that take advantage of both top-down and bottom-up strategies,
but also those which rely on texture analysis (Gabor analysis,
co-occurrence, HTD, edge histograms, etc.).

One critic that may be addressed to almost all of the existing
methods is their inability to process any image but binary ones, as
most of them generally need to separate the background from the
foreground of the document very early. The aim of the PhD is to design
a texture-based method, which would improve over the existing ones in
three different manners :

1. By using greylevel texture descriptors : the additional
information conveyed by greylevels should result in a significant
accuracy of the descriptors, and thereafter in that of the segmentation
itself. It would even be desirable to define or use color whenever
color is available.

2. By contextualizing the segmentation : although heterogeneous, the
document corpus remains rather well identified. Our idea is then to
introduce document models that could not only permit to modify the
probability laws that a pixel belongs to a class given the document and
the corpus, but also to give an a posteriori explanation of these laws
taken jointly. Bayesian networks, in particular, could constitute an
appealing framework to solve this problem.
3. By explicitly modelizing and explaining noise : it is highly
desirable that the segmentation method modelize noise as a class of its
own, and be able to explain it. Such an approach has already been
proposed in the literature, for instance for distinguishing between
handwritten and printed text on binary images by Zheng et al., and
exhibited interesting results. Significant improvements are to be
expected by a similar approach extended in grey levels.


Be Sociable, Share!

No Comments

Post your comment comment

You must be logged in to post a comment.