doi:10.1016/S0031-3203(00)00127-8 
Copyright © 2001 Pattern Recognition Society. Published by Elsevier Science B.V.
Adaptive graphical pattern recognition for the classification of company logos
Dipartimento di Ingegneria dell'Informazione, Università di Siena, Via Roma 56, 53100 Siena, Italy
Received 9 March 2000;
revised 16 August 2000;
accepted 16 August 2000
Available online 6 July 2001.
References and further reading may be available for this article. To view references and further reading you must
purchase this article.
Abstract
When dealing with a pattern recognition task two major issues must be faced: firstly, a feature extraction technique has to be applied to extract useful representations of the objects to be recognized; secondly, a classification algorithm must be devised in order to produce a class hypothesis once a pattern representation is given. Adaptive graphical pattern recognition is proposed as a new approach to face these two issues when neither a purely symbolic nor a purely sub-symbolic representation seems adequate for the patterns. This approach is based on appropriate structured representations of patterns which are, subsequently, processed by recursive neural networks, that can be trained to perform the given classification task using connectionist-based learning algorithms. In the proposed framework, the joint role of the structured representation and learning makes it possible to face tasks in which input patterns are affected by many different sources of noise. We report some results that show how the proposed scheme can produce a very promising performance for the classification of company logos corrupted by noise.
Author Keywords: Artificial neural networks; Adaptive processing of data structures; Structured representation of graphical items; Contour tree algorithm; Classification of company logos
Fig. 1. Graphical representations by means of the contour tree algorithm. (A) The original image once the contours are detected; (B) the DOAG that is extracted from the image; (C) a noisy image and the detection of the approximate-inclusion between two contours i
c; (D) the DOAG that is extracted from the noisy image.
Fig. 2. Graphical representations based on the mutual position of the segmented areas. (A) Undirected graph expressing the adjacency of the regions. (B) Directed ordered graph obtained once a point of view is chosen.
Fig. 3. Construction of the encoding network corresponding to a recursive network and a DOAG. Note that o=3 (graph outdegree), and that the nil pointers are represented by proper frontier (initial) states. (A) The recursive model stating the dependencies among the variables. (B) The implementation of the model with a neural network. (C) The input DOAG. (D) The encoding network obtained by the unfolding of the neural network on the input graph.
Fig. 4. Some examples of noise produced by fax machines and Xerox copiers (top) and the corresponding images (bottom) obtained with the noise models adopted in the paper.
Fig. 5. Filtering and contour-tree extraction for three different types of noise. (a) The logo is corrupted only by salt and pepper noise that is effectively removed by filtering the image before contour detection. (b) The logo is additionally corrupted by a single stripe. (c) The logo is corrupted by salt and pepper, stripe and blob noise.
Fig. 6. The collection of logos which has been used to carry out the second group of experiments (test 2).
Fig. 7. Comparison between the average recognition rates of recursive neural networks and MLPs for increasing dimensions of the spots on the image.
Table 1. Recognition rate for different types and levels of noise
Table 2. Comparison between different classifiers with respect to invariance to rotations
Table 3. Average recognition rate on the 770 classifiers that discriminate between two classes
Abstract
Purchase PDF (468 K)
E-mail Article
Add to my Quick Links
Cited By in Scopus (7)
Purchase PDF (408 K)
, 
, M. Maggini and E. Martinelli
