Abstract: In this article, a system to detect rooms in architectural floor plan images is described. We first present a primitive extraction algorithm for line detection. It is based on an original coupling of classical Hough transform with image vectorization in order to perform robust and efficient line detection. We show how the lines that satisfy some graphical arrangements are combined into walls. We also present the way we detect some door hypothesis thanks to the extraction of arcs. Walls and door hypothesis are then used by our room segmentation strategy; it consists in recursively decomposing the image until getting nearly convex regions. The notion of convexity is difficult to quantify, and the selection of separation lines between regions can also be rough. We take advantage of knowledge associated to architectural floor plans in order to obtain mostly rectangular rooms. Qualitative and quantitative evaluations performed on a corpus of real documents show promising results.

Abstract: Sketch recognition is a discipline which has gained an increasing interest in the last
20 years. This is due to the appearance of new devices such as PDA, Tablet PC’s
or digital pen & paper protocols. From the wide range of sketched documents we
focus on those that represent structured documents such as: architectural floor-plans,
engineering drawing, UML diagrams, etc. To recognize and understand these kinds
of documents, first we have to recognize the different compounding symbols and then
we have to identify the relations between these elements. From the way that a sketch
is captured, there are two categories: on-line and off-line. On-line input modes refer
to draw directly on a PDA or a Tablet PC’s while off-line input modes refer to scan
a previously drawn sketch.
This thesis is an overlapping of three different areas on Computer Science: Pattern
Recognition, Document Analysis and Human-Computer Interaction. The aim of this
thesis is to interpret sketched documents independently on whether they are captured
on-line or off-line. For this reason, the proposed approach should contain the following
features. First, as we are working with sketches the elements present in our input
contain distortions. Second, as we would work in on-line or off-line input modes, the
order in the input of the primitives is indifferent. Finally, the proposed method should
be applied in real scenarios, its response time must be slow.
To interpret a sketched document we propose a syntactic approach. A syntactic
approach is composed of two correlated components: a grammar and a parser. The
grammar allows describing the different elements on the document as well as their
relations. The parser, given a document checks whether it belongs to the language
generated by the grammar or not. Thus, the grammar should be able to cope with
the distortions appearing on the instances of the elements. Moreover, it would be
necessary to define a symbol independently of the order of their primitives. Concerning to the parser when analyzing 2D sentences, it does not assume an order in the
primitives. Then, at each new primitive in the input, the parser searches among the
previous analyzed symbols candidates to produce a valid reduction.
Taking into account these features, we have proposed a grammar based on Adjacency Grammars. This kind of grammars defines their productions as a multiset
of symbols rather than a list. This allows describing a symbol without an order in
their components. To cope with distortion we have proposed a distortion model.
This distortion model is an attributed estimated over the constraints of the grammar and passed through the productions. This measure gives an idea on how far is the
symbol from its ideal model. In addition to the distortion on the constraints other
distortions appear when working with sketches. These distortions are: overtracing,
overlapping, gaps or spurious strokes. Some grammatical productions have been defined to cope with these errors. Concerning the recognition, we have proposed an
incremental parser with an indexation mechanism. Incremental parsers analyze the
input symbol by symbol given a response to the user when a primitive is analyzed.
This makes incremental parser suitable to work in on-line as well as off-line input
modes. The parser has been adapted with an indexation mechanism based on a spatial division. This indexation mechanism allows setting the primitives in the space
and reducing the search to a neighbourhood.
A third contribution is a grammatical inference algorithm. This method given a
set of symbols captures the production describing it. In the field of formal languages,
different approaches has been proposed but in the graphical domain not so much work
is done in this field. The proposed method is able to capture the production from
a set of symbol although they are drawn in different order. A matching step based
on the Haussdorff distance and the Hungarian method has been proposed to match
the primitives of the different symbols. In addition the proposed approach is able to
capture the variability in the parameters of the constraints.
From the experimental results, we may conclude that we have proposed a robust
approach to describe and recognize sketches. Moreover, the addition of new symbols
to the alphabet is not restricted to an expert. Finally, the proposed approach has
been used in two real scenarios obtaining a good performance.