Automatic, Parameter-Free Shape Recognition System

Award winner at DAGM Symposium Mustererkennung (Pattern Recognition) in Stuttgart, October 1998

We offer a system that is able to recognize, accurately locate, and inspect objects in instances that vary in size and/or orientation, even when their appearance is degraded and they are partially occluded.

• Entirely new method of pattern finding that fundamentally differs from the methods actually used in machine vision systems.
• Objects are represented and analyzed as geometric shapes as opposed to the most of standard methods that use vectors of shape parameters or pixel grids of grey values.
• New capabilities such as parameter-free identification of complex patterns, contour deformations, and missing parts without any restrictions on shape complexity, orientation, and scaling.
• Given an object contour, the system recognizes parameter-free the most similar object in the database of object contours
• The database can be automatically generated from object contours
• There is no training or parameter adjustment, only one image for one object is necessary
• A new developed prefilter, the shape abstraction, guarantees robustness to noise distortion

• The system is based on a novel, universal, and parameter-free shape similarity measure. Main features of this measure are:
  • The shape similarity measure is in accord with our visual perception. It is based on psychological results on human visual perception.
  • The shape similarity measure respects visual parts of objects:
    First the best correspondence of visual parts is computed, then the similarity of the corresponding parts is determined.
  • The shape similarity measure is universal, it can be applied to contours of any shape and does not require parameter adjustment before.

In contrary, parametric methods and methods based on vectors of shape parameters like area, perimeter, elongation (major axis / minor axis), etc. are not universal (e.g., for a given vector of parameters, there always exist two shapes that cannot be distinguished) and require parameter adjustment when applied to new shapes (parameter determination in kind and value is necessary).

The system recognizes the exact position of the object, e.g., rotation angle with respect to the object in the database, this is due to correct part correlation of the prototype object and the search instance. Similarity and part correlation can be recognized even on non-rigid object deformations.

The system identifies missing parts and contour deformations

The system is able to recognize rough differences as well as differences in fine details

The system is able to abstract automatically from distortions, e.g. caused by digitization and segmentation noise. This is due to a novel discrete curve evolution method, the {\em shape abstraction}, which is applied to the contour before the similarity measure.

The shape abstraction used as a prefilter is very robust to noise distortions

The shape abstraction

• reduces influence of noise
• simplifies the shape by removing {\em irrelevant} shape features without changing {\em intuitive relevant} shape features.
• is parameter-free, hence it allows the abstraction of noise-distorted natural images as well as computer generated objects without parameter adjustment

The actual implementations are in C++ and JAVA. Only integer arithmetics is used, therefore the system can easily be realized even in hardware.

For demonstrational purposes an image database is available.

If you are interested in the Shape-Recognition-System or in any kind of cooperation please contact:

PD Dr. Longin Jan Latecki / Dipl.-Math. Rolf Lakämper
Institut für Angewandte Mathematik, Universität Hamburg
Bundesstr. 55
20146 Hamburg

Phone:
040-4123-5156 (Latecki)
040-4123-6241 (Lakämper)

E-mail:
latecki@math.uni-hamburg.de
lakaemper@math.uni-hamburg.de