An Analysis of the Road Signs Classification Based on the Higher-Order Singular Value Decomposition of the Deformable Pattern Tensors - Advanced Concepts for Intelligent Vision Systems - page 191

Application of tensors opened new possibilities for more precise analysis of complex
data which depend on many different factors. Each such factor is represented by a new
dimension of the tensor space (a mode of a tensor). In image processing different factors
correspond to different viewpoints, illumination conditions, or geometric deformations
of represented objects. This constitutes a qualitative difference compared to the matrix
approach in which images characteristic of different viewing conditions had to be
vectorized prior to the analysis, such as PCA [18]. Such tensor based methods have
been already used for CV tasks as handwritten character classification [19] or face
recognition [20], etc.
Specifically, in this paper we address the problem of reliable classification of the
road signs (RS) based on their monochrome pictograms. In the aforementioned
multilinear recognition framework, the task of RS classification is done with help of
the Higher-Order Singular Value Decomposition (HOSVD, called also the Tucker
decomposition) of the tensors built from the deformable versions of the prototype
patterns of each of the pictograms. To the best of our knowledge, this is the first
application of the HOSVD to the RS classification task. Nevertheless, as alluded to
previously, the presented framework can be also used for recognition of another group
of rigid objects, such as moving cars or fruits on a production line.
The work builds into our framework of RS recognition in which different detection
and classification modules were reported in [6-9]. In the group of developed classifiers
the presented in this paper tensor based method allows the best accuracy at very high
speed of response and manageable occupation of memory. More information pertinent
to the RS recognition task can be found in the works of de Escalera et al. [12], Paclik et
al. [17], Chen et al. [4], or Bascón et al. [3], etc., as well as in the mentioned references
[6-9].
The rest of the paper is organized as follows. We start with a discussion of the
architecture of the system. Then details of the tensor representation of patterns and
classification with the HOSVD are discussed. Further we discuss the implementation
issues related to the object-oriented computer representation of tensors as well as to
their so called proxy objects which allow efficient index manipulations of tensors
without data copying. Finally, we present the experimental results and conclusions.
2 Architecture of the Road Signs Recognition System
Architecture of our object recognition framework applied the task of RS recognition
is depicted in Fig. 1. It was designed to fit into our software framework developed
during the recent years [6-9]. However, in this paper we focus mostly on the HOSVD
based classification applied to the prohibition signs.
The preprocessing starts with the detection module which accepts an input color
image and returns rectangular outlines of the compact red objects, as described in [9].
Such rectangles are then cropped and then their color signal is converted into a
monochrome version by taking only the blue channel. Such a strategy showed to
provide the best contrast of the pictograms of the prohibition signs. Then, the detected
rectangle is registered to the size expected by the classification module, described in
the next section, as described in [9].