You are on page 1of 1

Cattle Brand Image Retrieval

Waldemar Villamayor Venialbo <wvilla@cpia.org> Centro Paraguayo de Informaciones Astronmicas Horacio Legal Ayala <hlegal@pol.una.py> Facultad Politcnica Departamento de Investigacin y Postgrado Universidad Nacional de Asuncin The aim of this project is to develop a method to detect similar logos in a cattle brand database, preventing the registration of new brands that could be confused with existing ones. The cattle brands we deal with are sketches handwritten with a digital pen within a limited rectangular region. These sketches are composed by one or more isolated draws: letters, numbers, symbols, etc.; many times these symbols appears overlapped. The proposed method is divided in four stages: 1) 2) 3) 4) image preparation; stroke segmentation; feature extraction; and image retrieval. queried at the retrieval stage. The query and database images are 8-bit grayscale images (256 gray levels). Brand sketches are composed of dark curves of the same thickness on a clear background. To have a workable image we do some normalization during the preparation stage, which consist in the binarization of the images resulting plain black and white bitmaps. Then we proceed segmenting the strokes contained in the images. At this stage we segregate each continuous curve that composes the brand logo. The segmentation method is an implementation of the orientation space projection technique described in [2] and [1]. The feature extraction stage is done using an implementation of the spatial multi-scale autoconvolution (MSA) method described in [3] and [4], which provides numerical descriptors for some combination of the components in the logo brand; such combination is done providing that the spatial distribution of the component strokes is not affected. These descriptorswhich are invariant to affine transformationare then interpreted in the retrieval stage to determine a similarity measure between two brand images. Such measure is obtained in a way that not only complete similar brands but subsets of a brand could be detected.

The first three steps applies to query and database images; the pre-processing of database images may be done a priori to populate a knowledge base that will be

Cattle Brand Retrieval Process


Set of features of query image Q

SQ =
Binarization Q1 Q2

MSA(Qi + Qj)
i=1 j=i+1

Stage 3
Select any databases image P for which

Q3

Q4

SQ
Qn

SP

Stage 1 References
[1] Cao, R. & Tan, C. L.; Line Primitive Extraction by Interpretation of Line Continuation. School of Computing, National University of Singapore. 2004. [2] Chen, J.; Sato, Y. & Tamura, S.; Orientation Space Filtering for Multiple Orientation Line Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(5), pp. 417429. 2000.

Stage 2

Stage 4

[3] Rahtu, E.; Salo, M. & Heikkil J.; Affine invariant pattern recognition using multi-scale autoconvolution. IEEE Transactions on Pattern Analysis and Machine Intelligence 27(6), pp. 908918. 2005. [4] Rahtu, E.; Salo, M. & Heikkil J.; A new efficient method for producing global affine invariants. Proc. International Conference on Image Analysis and Processing. pp. 407414. Springer. 2005.

You might also like