Chaur-Chin Chen

Bootstrap Techniques for Error Estimation

The design of a pattern recognition system requires careful attention to error estimation. The error rate is the most important descriptor of a classifiers performance. The commonly used estimates of error rate are based on the holdout method, the resubstitution method, and the leave-one-out method. All suffer either from large bias or large variance and their sample distributions are not known. Bootstrapping refers to a class of procedures that resample given data by computer. It permits determining the statistical properties of an estimator when very little is known about the underlying distribution and no additional samples are available. Since its publication in the last decade, the bootstrap technique has been successfully applied to many statistical estimations and inference problems. However, it has not been exploited in the design of pattern recognition systems. We report results on the application of several bootstrap techniques in estimating the error rate of 1-NN and quadratic classifiers. Our experiments show that, in most cases, the confidence interval of a bootstrap estimator of classification error is smaller than that of the leave-one-out estimator. The error of 1-NN, quadratic, and Fisher classifiers are estimated for several real data sets.

Improved moment invariants for shape discrimination

Abstract Moment invariants have been frequently used as features for shape recognition. They are computed based on the information provided by both the shape boundary and its interior region. Although several fast algorithms for computing traditional moment invariants have been proposed, none has ever shown the theoretical results of moment invariants computed based on the shape boundary only. This paper proposes improved moment invariants computed using the shape boundary only, which tremendously reduces computations. The new moment invariants, called improved moment invariants are mathematically proved to be invariant to scaling, translation, and rotation. Graphical plots of the first two improved moment invariants for six country maps and four industrial tools using improved moment invariants are also given. The results suggest that improved moment invariants can be used as effective features for shape discrimination or recognition.

MRF model-based algorithms for image segmentation

The authors empirically compare three algorithms for segmenting simple, noisy images: simulated annealing (SA), iterated conditional modes (ICM), and maximizer of the posterior marginals (MPM). All use Markov random field (MRF) models to include prior contextual information. The comparison is based on artificial binary images which are degraded by Gaussian noise. Robustness is tested with correlated noise and with object and background textured. The ICM algorithm is evaluated when the degradation and model parameters must be estimated, in both supervised and unsupervised modes and on two real images. The results are assessed by visual inspection and through a numerical criterion. It is concluded that contextual information from MRF models improves segmentation when the number of categories and the degradation model are known and that parameters can be effectively estimated. None of the three algorithms is consistently best, but the ICM algorithm is the most robust. The energy of the a posteriori distribution is not always minimized at the best segmentation.<<ETX>>

Color images' segmentation using scale space filter and markov random field

A new hybrid method is presented that combines the scale space filter (SSF) and Markov random field (MRF) for color image segmentation. The fundamental idea of the SSF is to use the convolution of Gaussian functions and image-histogram to generate a scale space image and then find the proper interval bounded by the local extrema of the derivatives. The Gaussian function is with zero mean and varied standard deviation. Using the SSF the different scaled histogram is separated into intervals corresponding to peaks and valleys. The MRF makes use of the property that each pixel in an image has some relationship with other pixels. The basic construction of an MRF is a joint probability given the original data. The original data is the image that is obtained from the source and the result is called the label image. Because the MRF needs a number of segments before it converges to the global minimum, the SSF is exploited to do coarse segmentation (CS) and then MRF is used to do fine segmentation (FS) of the images. Basically, the former is histogram-based segmentation, whereas the latter is neighborhood-based segmentation. Finally, experimental results obtained from using SSF alone, MRF using iterated conditional mode (ICM), and MRF using Gibbs sampling are compared.

Filtering methods for texture discrimination

Abstract Filtering methods have recently raised increasing interests in texture analysis due to their simulation of human vision. The goal of this paper is to evaluate the performance of four filtering methods including Fourier transform, spatial filter, Gabor filter and wavelet transform for texture discrimination. Experimental results on both natural textures and synthesized Markov random field (MRF) textures indicate that the wavelet features achieve almost the same recognition rate with the Gabor features, which is higher than the other two methods, whereas the computation time shows that the wavelet features are preferred.

Similarity measurement between images

Experimental results of applying two similarity measurements, Euclidean distance and chord distance, to test a set of six Brodatzs textures are reported. Experiments show that in addition to feature extraction, a similarity measurement between images should be simultaneously considered. We also review some other similarity measurements.

On the selection of image compression algorithms

This paper attempts to give a recipe for selecting one of the popular image compression algorithms based on: 1) wavelet, 2) JPEG/DCT, 3) vector quantisation, and 4) fractal approaches. We review and discuss the advantages and disadvantages of these algorithms for compressing gray-scale images, give an experimental comparison on four 256/spl times/256 commonly used images (Jet, Lenna, Mandrill, Peppers, and one 400/spl times/400 fingerprint image). Our experiments show that all of the four approaches perform satisfactorily when the 0.5 bits per pixel (bpp) is desired. However, for a low bit rate compression like 0.25 bpp or lower, the embedded zerotree wavelet approach and DCT-based JPEG approach are more practical.

EBARDenovo: highly accurate de novo assembly of RNA-Seq with efficient chimera-detection

MOTIVATION High-accuracy de novo assembly of the short sequencing reads from RNA-Seq technology is very challenging. We introduce a de novo assembly algorithm, EBARDenovo, which stands for Extension, Bridging And Repeat-sensing Denovo. This algorithm uses an efficient chimera-detection function to abrogate the effect of aberrant chimeric reads in RNA-Seq data. RESULTS EBARDenovo resolves the complications of RNA-Seq assembly arising from sequencing errors, repetitive sequences and aberrant chimeric amplicons. In a series of assembly experiments, our algorithm is the most accurate among the examined programs, including de Bruijn graph assemblers, Trinity and Oases. AVAILABILITY AND IMPLEMENTATION EBARDenovo is available at http://ebardenovo.sourceforge.net/. This software package (with patent pending) is free of charge for academic use only. SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.

SVD-based projection for face recognition

Projection-based face recognition has been widely studied during the past two decades. One of the problems is to require a huge storage space to save the face features obtained from training faces. This paper proposes an SVD-based face retrieval system which requires less memory than the PCA, 2DPCA, Fisher, and 2DFisher approaches. The algorithm tested on the famous ORL (AT&T) face database, consisted of 400 112 x 92 gray level face images equally contributed by 40 subjects, achieves 97.5% recognition rate of retrievals.

Markov random fields for texture classification

Abstract Texture features obtained by fitting generalized Ising, auto-binomial, and Gaussian Markov random fields (MRFs) to homogeneous textures are evaluated and compared by visual examination and by standard pattern recognition methodology. The MRF model parameters capture the strong cues for human perception, such as directionality, coarseness, and/or contrast. This paper is a comparative study of MRF model-based features. A comparison of classifying natural textures and sandpaper textures using nearest neighbor (NN), quadratic, and Fisher classifiers, suggests that both texture feature extraction and classifier design should be simultaneously considered in designing an optimal texture classification system.