Selecting the toroidal self-organizing feature maps (TSOFM) best organized to object recognition

Abstract- Se1f-organizing feature maps (SOFM) are an important tool to visualize high-dimensional data as a two-dimensional image. One of the possible applications of this network is in image recognition. However, this architecture presents some problems mainly due to border effects. In this paper a new organization of the feature maps termed toroidal self-organizing feature maps (TSOFM) is presented. Its main advantage consists in the elimination of the border effects and, consequent1y, an increase in the recognition rate. Another important aspect presented in this paper is the measurement of how well networks are organized during the training phase. This proposal has been experimented with a real data set, consisting of different pieces as resu1t to cut up chickens, demonstrating the better quality results. Several works [4], [5] propose criteria to reduce these effects. In this paper, we propose a new toroidal topology of the nodes for the SOFM ca11ed toroidal self-organizing feature map (TSOFM) and a degree of organization. In this new network architecture, all nodes are inner nodes and, consequently, there are no border effects in the network. The degree of organization is a parameter that permit us to know when a TSOFM is well organized avoiding over-training to obtain the best organized TSOFM. In section 2 we wil1 introduce the notation and some topological parameters. The next section discusses the border effects that appear in the SOFM. In section 5 we define a SOFM well organized and the degree of organization parameter. We apply this proposal to the recognition of different chicken pieces, and the experiments and results are described in section 6. Finally. some conclusion will be enumerated. I