Logo Recognition by Recursive Neural Networks
E.Francesconi P.Frasconi M.Gori
S.Marinai J.Q.Sheng G.Soda A.Sperduti
Dipartimento di Sistemi e Informatica-Universit`a di Firenze-Italy
Dipartimento di Ingegneria dell’Informazione-Universit`a di Siena-Italy Dipartimento di Informatica-Universit`a di Pisa-Italy
In this paper we propo an adaptive model,referred to as Recursive Neural Networks
(RRNNs)for logo recognition by explicitly conveying logo item into-ary tree repren-
tation,where symbolic and sub-symbolic information coexist.Each node in the contour-
tree is associated with an exterior or interior contour extracted from the logo instance.A
feature vector,which includes the perimeter of the contour,the area surrounded and the
印刷方式number of critical points at some pre-determined intervals,is associated with each node.
鼻涕带血The pattern reprentation transformed in this way contains the topological structured in-
formation of logo and continuous values pertaining to each contour node.Afterwards,the
RRNNs are ud to learn the logo regularities expresd by contour-trees.The experimen-
tal results are reported on40real logos with very promising results.
1Introduction
The adaptive computational scheme of statistical models and artificial neural networks is very well-suited for pattern recognition.It is commonly recognized,however,that the adap-tive models,especially neural networks,can hardly deal with highly-structuredinformation.On the other hand,most of the models propod in thefield of syntactic pattern recognition are very-well suited for incorporating pattern grammatical regularities,but,one of their drawbacks is that they are not strongly oriented to face the prence of noi.This scenario ems to be mainly due to the development of computational approaches that are basically oriented to either dealing with symbols or with unstructured information,whilst in thefield of pattern recognition one is often looking for a balanced combination of the symbolic and sub-symbolic processing levels.This paper is intended to bridge the
two distinct paradigms by applying RRNNs to solving logo recognition problems,a typical task in which the significant prence of noi and the structured information associated with the patterns might require both paradigms be taken into account.
Logo recognition has been investigated extensively as an important pattern recognition task in the last few years.Many of the propod approaches are bad on syntactic approach[1]and statistical model([2][3]).A connectionist-bad approach to logo recognition,has been recently propod in[4]and extended to deal with spot nois[5].In the papers,the logo membership is estimated by multilayer perceptrons acting as autoassociatorsthat are fed by a vector of unstructured features.Such an approach relies completely on the suppod capabil-ities of multilayer perceptrons to learn the logo regularities and face the prence of noi.Al-though very satisfactory results have been achieved,the effectiveness of this approach is likely to decrea for pattern recognition applications in which there is a strong symbol’s component.
On the other hand,in order to unify symbolic and sub-symbolic processing,instead of solely
processing vector of real values,some rearchers have recently investigated the feasibility of learning from examples structured data by using recursive artificial neural networks(RRNNs) [6].The 关雎蒹葭
propod neural model can be straightforwardlyud for solving logo recognition prob-lem,while the noi on the logo instance and the structured information derived from logos can be properly incorporated into RRNNs.We propo to make the logo structure explicit by offer-ing the neural networks a reprentation which is bad on tree structures,obtained from logos by using a recent algorithm propod by Cortelazzo et al.[2].An-ary tree is created by re-cursively detecting contours and inclusion relations,so that nodes in odd levels of the tree are associated to exterior contours and even levels are associated to interior contours.Each node is then labeled by a t of numerical features measured on the object associated to the node and collected in a multivariate variable.Examples of the features are the perimeter of the contour,the area surrounded by the contour,and the number of critical points at some predeter-mined intervals.Scale invariance is obtained by appropriate normalization of continuous label and rotation invariance is gained by appropriate ordering of the children for a given node.
The paper is organized as follows:in Section2,the tree reprentation of logos is prented while in Section2.3we describe our approaches of tree pruning and grafting.In Section3we describe with more detail the RRNN model and its relevance for logo recognition;In Section4 we briefly introduce the network architecture ud for our experiments for the logo recognition. In Section5we report the experimental results on40logos andfinally some conclusions are drawn in Section6.
2Tree Construction
Structural reprentations of patterns can be gained by region-and contour-bad approaches. The region-bad approach derives afixed number of out-degree of tree and contour-bad ap-proach constructs a tree by detecting contours and their relationship.We have chon a recent algorithm propod by by Cortelazzo et al.[2]to construct tree reprentation of logos.The algorithm creates an-ary tree by recursively detecting contours and inclusion relations(e examples in Fig.3),so that nodes in odd levels of the tree are associated to exterior contours and even levels are associated to interior contours.In this paper,the tree obtained by means of the algorithm is called primary contour-tree.It reprents primary contour items and the topo-logical inclusion relationships among extracted contours.Although the primary tree could be undertaken a graph pruning and grafting step in order to avoid learning very high valence tree, the tree nodes reprenting significant primitives remain unchanged,main features pertaining to the logo instance are reasonably retained.When a primary contour-tree be pruned and grafted,a condary contour-tree is generated by deleting some nodes and links from the primary contour-tree and adding new nodes to the tree.The motivations for pruning and grafting primary tree are described in Section2.3in great detail.
Obviously,each node in either primary or condary contour-tree is labeled by a t of nu-merical fea
tures measured on the object(s)associated to the node(s),collected in a multivariate variable.Appropriate normalization of continuous labels assures the scale-invariance and rotation invariance can be obtained through an ordering operation of children for a given node. In Fig.2we depict how logo can be reprented by contour tree and their recursive neural mod-els.By investigating the examples,it was noted that the feature vector remains unchanged after normalization of contour perimeters and areas,thus assuring the scale invariance.Furthermore, the ordering of the children according to contour topological characteristics for a node can guar-antee the invariance of tree architecture,conquently,rotation-invariance is warranted,other-wi,the order of the children for a given node can be interchanged since the contour-tracing can be operated stochastically in the ca of logo rotation.
2.1Contour Tracing
Contour tracing is a local operation in a binarized image by traversing to one8-neighbor pixel from a black one(suppo the white is background color).Contour tracing and contour-tree construction follows the algorithm described in[2].We have ud incoming direction for contour pixels in order to memorize the Freeman code instead of traditional outgoing direction. Pixel in the orthogonal the incoming direction in clockwi order isfirst visited when scanning the raster image
from the upper-left corner to bottom-right corner.The exterior contour is formed with clockwi ordered pixels and interior contour with anti-clockwi or-dered pixels.Perimeter of contour is obtained directly in the pha of contour tracing by count-ing the even direction pixel as1and odd direction pixel as
from the logo instance.We denote them using small letters as follows:
Pruning node node)
2if Outdegree(node)V ALENCE then
I nitialize values No big=Small[]=Big[]=0七夕风俗
S earch small and big children for the node
3for Outdegree(node)do
C heck the dimension of th associated contour of node
4if th node is BIG then
5Big[No
small++]
8if No
node
10Search all children of small nodes 11if Exist children then
12Create a new Tree
big No
big V ALENCE then
18No node No V ALENCE-1 19for do
20Create a new Tree
Pruning
Pruning
我们将战斗到底of the backpropagation algorithm is available for computing the gradient by error propagation through structure
旭日什么什么成语[6].
我学会了骑自行车
Contour Trees Recursive Models
Logo Images
exterior contour 212e 4e 0
e Figure 2:Two logo images are converted to the corresponding tree reprentations and to their recursive models.Squares indicate null pointer and black circle indicate actual value vectors associated with the corresponding contour.
4Network Architecture
交友的句子A MLP-like network architecture with one hidden layer is employed to realize the RRNNs in logo recognition.
Input Layer
Figure 3:The MLP-like architecture employed for logo recognition
From the figure 3,it is shown that the output neurons are fully connected with the hid-den neurons and state neurons are ud to memorize the state vectors.In the learning process,the state vector is copied to the input layer according to the topological structure of condary contour-trees .
