Chengan Guo

An SVM classification algorithm with error correction ability applied to face recognition

This paper presents an SVM classification algorithm with predesigned error correction ability by incorporating the error control coding schemes used in digital communications into the classification algorithm. The algorithm is applied to face recognition problems in the paper. Simulation experiments are conducted for different SVM-based classification algorithms using both PCA and Fisherface features as input vectors respectively to represent the images with dimensional reduction, and performance analysis is made among different approaches. Experiment results show that the error correction SVM classifier of the paper outperforms other commonly used SVM-based classifiers both in recognition rate and error tolerance.

A vector quantization approach for image segmentation based on SOM neural network

In the existing segmentation algorithms, most of them take single pixel as processing unit and segment an image mainly based on the gray value information of the image pixels. However, the spatially structural information between pixels provides even more important information of the image. In order to effectively exploit both the gray value and the spatial information of pixels, this paper proposes an image segmentation method based on Vector Quantization (VQ) technique. In the method, the image to be segmented is divided into small sub-blocks with each sub-block constituting a feature vector. Further, the vectors are classified through vector quantization. In addition, the self-organizing map (SOM) neural network is proposed for realizing the VQ algorithm adaptively. Simulation experiments and comparison studies have been conducted with applications to medical image processing in the paper, and the results validate the effectiveness of the proposed method.

Face recognition based on gabor-enhanced manifold learning and SVM

Recently proposed Marginal Fisher Analysis (MFA), as one of the manifold learning methods, has obtained better classification results than the conventional subspace analysis methods and other manifold learning algorithms such as ISOMAP and LLE, because of its ability to find the intrinsic structure of data space and its nature of supervised learning as well In this paper, we first propose a Gabor-based Marginal Fisher Analysis (GMFA) approach for face feature extraction, which combines MFA with Gabor filtering The GMFA method, which is robust to variations of illumination and facial expression, applies the MFA to augmented Gabor feature vectors derived from the Gabor wavelet representation of face images Then, the GMFA method is integrated with the Error Correction SVM classifier to form a novel face recognition system We performed comparative experiments of various face recognition approaches on ORL database and FERET database Experimental results show superiority of the GMFA features and the new recognition system presented in the paper.

A Parallel Image Segmentation Method Based on SOM and GPU with Application to MRI Image Processing

This paper presents a parallel image segmentation method based on self-organizing map (SOM) neural network by extending the authors’ former work from serial computation to parallel processing in order to accelerate the computation process. The parallel algorithm is composed of a group of parallel sub-algorithms for implementing the entire segmentation process, including parallel classification of the image into edge/non-edge pattern vectors, parallel training of an SOM network, and parallelly segmenting the image by using the trained SOM model with vector quantization approach. In the paper, the parallel algorithm is implemented on GPU with OpenCL program language and applied to segmenting the human brain MRI images. The experimental results obtained in the work showed that, compared with the original serial algorithm, the parallel algorithm can achieve a significant improvement on the computation efficiency with a speedup ratio of 64.72.

A Parallel Implementation of Error Correction SVM with Applications to Face Recognition

The Error Correction SVM method is an excellent multiclass classification approach and has been applied to face recognition successfully. Yet, it suffers from the computational complexity. To reduce the computation time of the algorithm, a parallel implementation scheme is presented in the paper in which the training and classification tasks are assigned to multiple processors and run on all the processors simultaneously. The simulation experiments conducted on a local area network using Cambridge ORL face database show that the parallel algorithm given in the paper is effective in speeding up the algorithms of the training and classification while maintaining the recognition accuracy unchanged.

An Optimal Neural-Network Model for Learning Posterior Probability Functions from Observations

This paper presents the further results of the authors’ former work [1] in which a neural-network method was proposed for sequential detection with similar performance as the optimal sequential probability ratio tests (SPRT) [2]. The analytical results presented in the paper show that the neural network is an optimal model for learning the posterior conditional probability functions, with arbitrarily small error, from the sequential observation data under the condition in which the prior probability density functions about the observation sources are not provided by the observation environment.

Face recognition based on sparse representation and error correction SVM

Very recently, the sparse representation theory in pattern recognition has aroused widespread concern. It shows that a sample can be linearly recovered by the others in the database and the coefficients are sparse. Based on this theory, this paper proposed a new feature extraction algorithm-Sparse Representation Discrimination Analysis (SRDA) by combining the sparse representation theory and the manifold learning model together. The SRDA algorithm can maintain not only the sparse reconstruction relationship of original data, but also the spatial structure in low dimensional space. Then, the SRDA feature is integrated with the error correction SVM to build a new face recognition system. Comparative experiments of various face recognition approaches are conducted by testing on the ORL, AR and FERET databases in the paper and the experimental results show the superiority of the new method.

Face recognition based on gabor enhanced marginal fisher model and error correction SVM

In previous work, we proposed the Gabor manifold learning method for feature extraction in face recognition, which combines Gabor filtering with Marginal Fisher Analysis (MFA), and obtained better classification result than conventional subspace analysis methods. In this paper we propose an Enhanced Marginal Fisher Model (EMFM), to improve the performance by selecting eigenvalues in standard MFA procedure, and further combine Gabor filtering and EMFM as Gabor-based Enhanced Marginal Fisher Model (GEMFM) for feature extraction. The GEMFM method has better generalization ability for testing data, and therefore is more capable for the task of feature extraction in face recognition. Then, the GEMFM method is integrated with the error correction SVM classifier to form a new face recognition system. We performed comparative experiments of various face recognition approaches on the ORL, AR and FERET databases. Experimental results show the superiority of the GEMFM features and the new recognition system.

A Two-Pass Classification Method Based on Hyper-Ellipsoid Neural Networks and SVM's with Applications to Face Recognition

In this paper we propose a two-pass classification method and apply it to face recognitions. The method is obtained by integrating together two approaches, the hyper-ellipsoid neural networks (HENNs) and the SVMs with error correcting codes. This method realizes a classification operation in two passes: the first one is to get an intermediate classification result for an input sample by using the HENNs, and the second pass is followed by using the SVMs to re-classify the sample based on both the input data and the intermediate result. Simulations conducted in the paper for applications to face recognition showed that the two-pass method can maintain the advantages of both the HENNs and the SVMs while remedying their disadvantages. Compared with the HENNs and the SVMs, a significant improvement of recognition performance over them has been achieved by the new method.

Bird Species Recognition Based on Deep Learning and Decision Fusion

Bird species recognition is one of the most challenging tasks in fine-grained visual categorizations (FGVC) and has attracted wide attention in recent years. In this paper, we develop a bird recognition system that consists of three learning-type computational modules: the first one is for extracting the object and key parts from input images that is implemented by training a deep convolutional neural network (CNN) model, the second module is for feature extraction from the detected object and parts by using four other CNNs and further for computation of the high dimensional Gaussian descriptors based on the deep features, and the third module is for getting the final recognition result that is implemented by training four SVM classifiers with the Gaussian descriptors and integrating the outputs of the SVMs together with a decision fusion method. Experiment results obtained in the paper confirm the validity of the proposed approach.