Computer Vision Projects in Image and Pattern Analysis Group of SZTAKI
by Dmitry Chetverikov
Image and Pattern Analysis Research
Group (IPAN) belongs to the Geometrical Modelling Laboratory of SZTAKI.
As a unit of the Research Division at the Institute, the group does research
in various areas related to computer vision and its applications. The main
activities of IPAN are in texture, shape and motion analysis, including
a technologically oriented European project devoted to industrial quality
control of ferrite cores and a research project on query-by-image-content
in image databases.
Texture plays a key role both in human and machine vision. Textures
are non-figurative, repetitive patterns that appear on surfaces of materials,
in textiles, indoor, outdoor and aerial imagery (forests, fields, towns,
etc.), biological and other microscopic images, in documents, eventually
in all the categories of images. Basic, perceptually motivated properties
of textures are their regularity, anisotropy, and complexity. Recently,
a powerful approach to texture analysis, called feature based interaction
map (FBIM), has been developed in IPAN. FBIM has been successfully applied
in a number of important problems and tasks, including the evaluation of
the basic properties as well as texture symmetry; precise determination
of texture pattern orientation; rotation-invariant texture classification;
adaptive texture filtering selective to local structure and its orientation;
invariant detection of textured objects and texture defects; document zone
classification; estimation of document page skew for subsequent skew normalization.
Many of the FBIM based solutions are already supported by the results of
systematic experiments with large test image databases, such as the UW-I
document image database compiled at the University of Washington in Seattle.
A related research project will be started soon aimed at query-by-image-content
in image databases using keys as texture regularity and anisotropy as well
as structure orientation. The concept of FBIM based adaptive structural
filtering will be applied to rotation-, scale-, and surface tilt-invariant
retrieval of images that contain regions similar in structure to the query
pattern. Both texture and shape keys will be used, including characteristic
configurations of specific, readily detectable feature points such as corners.
Recently, a novel efficient algorithm for detection of corners and other
points of high contour curvature has been developed in IPAN.
Figure 1a-b: Detecting structural defect in texture.
Left: A textile texture with a structural defect.
Right: The detected defect overlaid on the texture.
Motion analysis is of growing relevance for computer vision since many
of real-world vision tasks involve dynamic, time-varying images. Last year,
a new research direction was opened in IPAN with the goal of creating methods
and algorithms for tracking independently moving, multiple small-size objects
represented by their feature points. The points may occur in groups, temporarily
disappear, and leave and enter the viewfield, which significantly complicates
the tracking problem. This study is based on a previously developed and
published algorithm for matching of two subsequent frames of a moving point
set assuming smooth trajectories and prior knowledge of the expected direction
of displacement for each point. In the current, more realistic formulation,
the tracking will be done for long motion sequences without any prior knowledge
concerning the directions of displacements. A motion generator has already
been developed that will provide a testbed for future tracking experiments.
Several alternative tracking algorithms have been implemented for comparative
testing, including the pilot version of the new IPAN algorithm.
Figure 2a-c: Examples of trajectories synthesized by the motion generator.
In the framework of the EU COPERNICUS research project CRASH (CRAck
and SHape defect detection in ferrite cores), an industrial machine vision
system has been developed with the goal of detecting and measuring surface
and shape defects in ferrite cores. Partners involved in CRASH project
are the University of Trento, Italy (project management, signal processing);
Technical Software Consultants Ltd., Great Britain (electromagnetic sensor
for detection of inner defects); POLFER Magnetic Materials Ltd., Poland
(task specification, system validation); IBIB Polish Academy of Sciences
and Association of Image Processing Inc., Poland (detection of surface
defects). IPAN group created an integrated system for shape measurements
of ferrite cores based on the basic projections (views) of the shapes.
The system has been successfully tested on a large set of ferrite core
images. Using an original subpixel approach to comparison of a measured
core shape to the ideal shape, results consistent with the reference manual
measurements at the production line have been obtained. Currently, this
project is being supported by another grant, INFO-COPERNICUS grant, with
the goal of putting CRASH results into industrial practice and facilitating
the development of related measurement standards.
Figure 3a-c: Three views of a defective E-shaped core. The shape defect
(bent leg) can be observed in the frontal projection.
For information on IPAN group, see web page at http://leader.ipan.sztaki.hu/.
The CRASH home page address is http://www.lii.unitn.it/CRASH/.
Please contact:
Dmitry Chetverikov - SZTAKI
Tel: +36 1 2096510
E-mail: csetverikov@sztaki.hu
