Xiaofang Tang

Comparison BAM and discrete Hopfield networks with CPN for processing of noisy data

In the paper we compared three neural networks -Koskopsilas the bidirectional associative memory (BAM) and the discrete Hopfield network (DHN) with the counter propagation network (CPN) for processing of noisy data. We probe into their commonness and distinctness. The experimental results show that de-noise results of three neural networks for weak noise are almost same. BAM of the gradient-descent algorithm is the best for de-noisy processing, at some condition Koskopsilas BAM network is of the same performance as the discrete Hopfield network which is better than the CPN for strong noise.

Implicit Surface Reconstruction from Noisy 3D Scattered Data

The paper presents a method for surface reconstruction from large unorganized and noisy point sets without any normal or orientation information. Firstly, the outliers will be selected and deleted, and acquire new point sets being less noisy with improved method of fuzzy c-means clustering. Secondly, we compute the normal of the new point sets through PCA analysis. Then we interpolate single-level and multi-level the new points acquired by clustering method by compactly supported radial basis function (CSRBF). The experiments show that the outliers will be selected well and the noisy point sets will be smoothed. The interpolation points are renewed by clustering method, and implicit reconstruction by CSRBF may provide a solution for surface reconstruction from noisy and incomplete data. In our algorithm, we represent the point sets by an octree-based subdivision of the bounding box to reduce computation complexity

Gabor feature based classification using 2d linear discriminant analysis for face recognition

This paper introduces a novel 2D Gabor-Fisher Classifier for face recognition. The 2D-GFC method applies the 2D Fisher Linear Discriminant Analysis (2D-LDA) to the gaborfaces which is derived from the Gabor wavelets representation of face images. In our method, Gabor wavelets first derive desirable facial features characterized by spatial frequency, spatial locality, and orientation selectivity to cope with the variations due to illumination and facial expression changes. 2D-LDA is then used to enhance the face recognition performance by maximizing the Fishers linear projection criterion. To evaluate the performance of 2D-GFC, experiments were conducted on FERET database with several other methods.

Background model initialization in moving object detection with shadow elimination

Background subtraction methods are widely exploited for moving object detection in many applications. How to correctly and efficiently model the background is the key technique to such approaches. We review the typical systems characterized by pixel-wise multiple Gaussians statistics. Our research on model initialization is introduced afterwards, where a fast online initialization algorithm is devised to train the model quickly and correctly. Then a convenient way to combine luminance distortion with chrominance distortion is proposed for shadow detection in complex scene. Finally, relevant experimental results are provided to highlight our methods.

A fast learning algorithm of feedforward neural networks by using novel error functions

This paper presents two novel alternative families of error functions as the generalized training criterion of feedforward neural networks; they can significantly accelerate the convergence rate in the midterm and the last training stages. Their training speed is faster than the original fast backpropagation algorithm by parameter optimization. Several approaches to parameter optimization are explored and verified by experiments.

Error correction via phonetic similarity-based processing for chinese spoken dialogue system

This paper reviewed the approaches of error handling in spoken dialogue systems in the literature at first. To take advantage of domain knowledge and the characteristics of the errors in Chinese speech recognition, a similarity between text strings from speech recognizer and the correct words was defined. A similarity-based algorithm of error handling in Chinese spoken language understanding was developed to improve the correctness and robustness of sentence understanding in spoken dialog systems. Before the procedure of keyword spotting, the errors in text strings are revised first. Preliminary experimental results show that the algorithm is effective

Wavelet neural networks for adaptive equalization

A structure based on the wavelet neural networks is proposed for nonlinear channel equalization in a digital communication system, the minimum error probability (MEP) is applied as performance criterion to update the weighting matrix of wavelet networks. Our experimental results show that performance of the proposed wavelet networks based on equalizer can significantly improve the neural modeling accuracy and outperform the conventional neural networks in signal to noise ratio and channel non-linearity.

3D Object Classification by Part Features Fusion

The paper proposes a part-level features fusion method for 3D object classification. 3D object is represented as a combination of its constituent parts, which are reconstructed by superquadric models. The presented classification method is decomposed into three main phrases: part features extraction of 3D objects, training set construction and classification implementation. Experimental results show the validity and potential of the presented method for part-level 3D object classification.

Superquadric-based 3D Scene Reconstruction and Interpretation

In this paper, an integrated processing framework is proposed for 3D scene reconstruction and interpretation. Superquadric-based 3D scene reconstruction from both 2D images and 3D data are developed, in which superquadric-based hierarchical description is implemented as the universal parametric representation of real 3D scene objects. A volumetric part based 3D object similarity match approach is developed for the recognition and classification in semantic interpretation of virtual scene. Experimental results are presented, which show the efficiency and potential advantages of the integrated framework in 3D scene reconstruction and interpretation for automatic reconstruction and interactive manipulation of virtual 3D scene

Neural Networks for Clustering Analysis of Molecular Data

In this paper hierarchical cluster and the competitive learning cluster are compared by using molecular data of large size sets. We construct a reproducible matrix to evaluate the quality of clustering, and dead nodes problem of the competitive learning network is solved by the conscience mechanism. The experimental results show that the hierarchical clustering can represent a multi-level hierarchy which show the tree relation of cluster distance, the competitive learning network has a good clustering reproducible and indicate the effectiveness of clusters for molecular data