Elsevier

Pattern Recognition Letters

Volume 23, Issue 12, October 2002, Pages 1449-1457
Pattern Recognition Letters

Integration of local and global shape analysis for logo classification

https://doi.org/10.1016/S0167-8655(02)00105-8Get rights and content

Abstract

A comparison is made of global and local methods for the shape analysis of logos in an image database. The qualities of the methods are judged by using the shape signatures to define a similarity metric on the logos. As representatives for the two classes of methods, we use the negative shape method which is based on local shape information and a wavelet-based method which makes use of global information. We apply both methods to images with different kinds of degradations and examine how a particular degradation highlights the strengths and shortcomings of each method. Finally, we use these results to develop a new adaptive weighting scheme which is based on the relative performances of the two methods. This scheme gives rise to a new method that is much more robust with respect to all degradations examined and works by automatically predicting if the negative shape or wavelet method is performing better.

Introduction

We examine three different approaches for classifying images with several components in an image database. One approach uses a local method to represent the image, the second uses a global method, while the third combines both using an adaptive weighting scheme based on relative performance. The local method uses so-called negative symbols, as described in Soffer and Samet (1998), to compute a number of statistical and morphological shape features for each connected component of an image foreground and background. The global method uses a wavelet decomposition of the horizontal and vertical projections of the global image as described in Jaisimha (1996). As a sample application of well-defined multi-component images, we use logos. Several studies have reported results on some form of logo recognition. Each study used either global or local methods. These include local invariants (Doermann et al., 1996; Kliot and Rivlin, 1997), wavelet features (Jaisimha, 1996), neural networks (Cesarini et al., 1997), and graphical distribution features (Kato, 1992). The performance in case of certain degradations was examined. In this paper we compare the local and global methods under the influence of several image degradations. The performance measure is the ranking of the original logo after inputing a degraded version of it into the classifier. The results exhibit the advantages and disadvantages of local methods, based on shape features, in contrast to global methods, rooted in signal processing. Finally, we present an algorithm that combines both methods into a single, robust framework by adaptively weighting the contributions of each method according to an estimate of their relative performance.

Section snippets

Wavelet method

We studied logos in the UMD-Logo-Database which are gray-scale images that are scanned versions of black and white logos (Fig. 1a). We assume that the logos have already been segmented and binarized by a preprocessing step. The problems of segmentation and threshold selection are beyond the scope of this paper. The classification methods should be scale, translation, and rotation invariant. To achieve this, we apply some normalizing steps to the input images before we start the computation of

The negative shape method

The novel idea of the negative shape method as defined in Soffer and Samet (1998) for the representation of symbol-like data such as found in logos is that we compute the shape features not just for the components of the foreground that constitute the symbol itself, but also for the components that make up the background of the image containing the symbol.

We start by labeling the connected components of the image and its background. To each component of the labeled image, we first apply the

Comparison between the methods

All methods were implemented in MatlabTM and were applied to the logos contained in the UMD-Logo-Database (123 logos, can be found at http://documents.cfar.umd.edu/resources/database/umdlogo.html). The system was tested by providing it with an input logo and ranking the logos in the database based on their similarity to this logo. Below, we investigate the robustness of the methods when the logos are corrupted using four different image degradation methods. We choose two global degradations as

The best of both worlds: a combination of the two methods

In Section 4 we saw that the wavelet and the negative shape methods perform very differently if the input logo is corrupted by either local or global degradations. Thus, if we were able to detect which method is performing better, then we could devise a switching mechanism that chooses the better performing method automatically.

To be able to compare both methods, we first normalize the featured distances by computing for both methods the average of the pairwise feature distances between all

Summary and future work

Both the wavelet as well as the negative shape method are well-suited for certain kinds of image degradations but are very sensitive to others. This discrepancy in performance can be explained by the difference between local shape feature-based methods and global filter-based methods. On the one hand, we have the wavelet method that operates on the global image and computes features that are relatively invariant to degradations that are isotropic. On the other hand, we have the negative shape

References (8)

  • Cesarini, F., Fracesconi, E., Gori, M., Marinai, S., Sheng, J.Q., Soda. G., 1997. A neural-based architecture for...
  • D Doermann et al.

    Applying algebraic and differential invariants for logo recognition

    Machine Vision and Applications

    (1996)
  • Jaisimha, M.Y., 1996. Wavelet features for similarity based retrieval of logo images. Proceedings of the SPIE, Document...
  • T Kanungo et al.

    Statistical nonparametric methodology for document degradation model validation

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2000)
There are more references available in the full text version of this article.

Cited by (50)

  • Logo and seal based administrative document image retrieval: A survey

    2016, Computer Science Review
    Citation Excerpt :

    Local features are summarized according to their extraction for each point of input domain, whereas global features are extracted based on sets of pixels, on a region or even on the whole document. Based on this categorization outline, local features used for logo recognition include: features extracted from local zone [54], differential invariants [5], negative shape features [6], primitives (line segments) [55], curvature and distance from centroid point [12,7], SIFT and SURF descriptors derived from Hessian-affine interest points [56–58], horizontal gaps per total area, vertical gaps per total area, ratio of hole area to total area [59,60], color [61], Delaunay triangulation of components/local features [61,60], bag-of-words features [60], edge based features extracted using GHT [62], Fourier coefficients of segmented boundary curves [63], rectangle features extracted from integral image [64], etc. Global features utilized in literature for logo recognition are: different moments (Zernike, Tchebichef, invariant, radial) [54,11,59,65,2,7], projection profiles [66], bispectral [66], gradient features extracted from contour points [67–69], algebraic invariants [5], wavelet-based features [6,70], circularity, eccentricity and rectangularity [59], geometric topology features extracted from components [61], area, isolation, deviation, symmetry, centralization, complexity and 2-level contour representation strings [71], global shape based features (circle, rectangle, triangle, ellipse, polygon, and B-spline) extracted from Fourier descriptor [62,72], shape context [73,3], features extracted from raw image/vector data [74], curvature [22], template matching [75], etc.

  • Sparse representation over learned dictionary for symbol recognition

    2016, Signal Processing
    Citation Excerpt :

    In general, these descriptors are invariant under similarity transformations, but it depends on a prior normalization step, which is very sensitive to noise. Although these descriptors can easily be computed, either they are usually poorly discriminant [41] or the matching process is time consuming [42]. Syntactic and structural descriptors are suitable for symbol description since differences between symbols come from differences between spatial relation between primitives (e.g. lines and arcs).

  • Features Fusion for Classification of Logos

    2016, Procedia Computer Science
View all citing articles on Scopus
1

The support of the National Science Foundation under Grants CDA-95-03994, IRI-97-12715, EIA-99-00268, and IIS-00-86162 is gratefully acknowledged.

2

Currently at IBM Research Lab, Haifa 31905, Israel.

View full text