Wyniki wyszukiwania - BazTech

1

Center selection of RBF neural network based on modified k-means algorithm with point symmetry distance measure

Sing J. K., Basu D. K., Nasipuri M., Kundu M.

Foundations of Computing and Decision Sciences

|

2004

|

Vol. 29, No. 3

247-266

EN

The hidden layer neurons of a radial basis function (RBF) neural network map input patterns from a nonlinearly separable space to a linearly separable space. To locate the centers of those hidden layer neurons, normally k-means clustering algorithm is used. Normal k-means clustering algorithm cannot detect hyper spherical-shaped clusters along the principal axes. In present study, we propose a modified version of the k-means clustering algorithm to select RBF centers, which can eliminate this drawback. In the proposed algorithm, we modify the k-means algorithm in two stages. In trie first stage, the procedure to select the initial cluster centers has been modified to capture more knowledge about the distribution of input patterns. In the second stage, the initial centers, selected in the first stage are updated using point symmetry distance measure instead of using conventional Euclidean distance. The RBF neural network with the proposed algorithm has been tested with three different machine-learning data sets. It has also been applied for the segmentation of medical images. The experimental results show that the RBF neural network using the proposed modified k-means algorithm performs better than that using normal k-means algorithm.

2

Integration of kernel and DCT coefficient fuzzy matching techniques for human face recognition

Bhattacharjee D., Basu D. K., Nasipuri M., Kundu M.

Foundations of Computing and Decision Sciences

|

2002

|

Vol. 27, No. 2

61-76

EN

In this paper we describe results of human face recognition using integration of two different fuzzy matching methods. The first method is based on the kernel matching concept. In this method the system is trained with different images of the same class. During training the presence of (55)-derived kernels are considered. Based on the presence or absence of kernels in different images of the same class one reference matrix is computed and using a fuzzy membership function on deviation of individual images from the values of reference matrix two tolerance matrices are created. In the second method using local shift invariant Discrete Cosine Transform (DCT) coefficients of images of a class, another set of reference matrix and two tolerance matrices are created in a same way as in the first method. For testing the image in a class, the values of the matching of input image with reference matrices for two different methods are calculated and classified based on their respective tolerance matrices. The final result is obtained by integrating the results from both the methods. The result of recognition is very high, where error rate is only 2.625% with false acceptance rate of 1.86%.

3

Robust texture classification using wavelet frames

Acharyya M., Kundu M.

Image Processing & Communications

|

1999

|

Vol. 5, no 2

19-37

EN

In this paper we present an approach to characterize textures at multiple scales using wavelet transforms and discuss the issues of translational and rotational invariance and noise immunity of a texture analysis system. We employ the non-separable discrete wavelet frames analysis which gives an overcomplete wavelet decomposition. Discrete Wavelet Frame (DWF) decompose the textures into a set of frequency channels. A texture is characterized by a set of these channel variances in this work. Classification experiments using twenty Brodazt textures indicate that texture signatures based on wavelet frame analysis are beneficial for accomplishing subtle discrimination of textures and robust classification against rotation translation and noise.