Catálogo de publicaciones - libros

A novel Support Vector Regression(SVR) algorithm has been proposed recently by us. This approach, called Lagrangian Support Vector Regression(LSVR), is an reformulation on the standard linear support vector regression, which leads to the minimization problem of an unconstrained differentiable convex function. During the process of computing, the inversion of matrix after incremented is solved based on the previous results, therefore it is not necessary to relearn the whole training set to reduce the computation process. In this paper, we implemented the LSVR and tested it on Mackey-Glass time series to compare the performances of different algorithms. According to the experiment results, we achieve a high-quality prediction about time series.

A new approach for maximum posterior probability (MAP) image restoration based on cellular neural network (CNN) is proposed in this paper, and hardware realization is also discussed. According to analysis of MAP image restoration, a new template is proposed for CNN image restoration. The computer simulation result proves the approach is reasonable, then a hardware system based on CNN processor is setup for the restoration algorithm, and the effectiveness of the CNN processor is also confirmed in this system.

The enlargement of the digital image implies the improvement of the image resolution, where the high frequency components lost in sampling must be estimated. In this paper, an image enlargement method using a high resolution neural network is proposed, corresponding to the region with rapid change (high local variance) and the region requiring a smooth interpolation (low local variance). It is shown that the high resolution NN has high potential ability.

A novel edge enhancement based on Hopfield neural net is presented in this paper, which is a post-processing complement for a pre-existing edge detector. This term is added to the output of the edge detector. Firstly, the energy function which is used to find the final stable edges is provided in the Hopfield neural net, and then, based on the window iteration, it improves the performance of the edge detector by recovering missing edges and eliminating false edges. In experiments conducted on various images, we demonstrate the performance of the algorithm on them.

A quasi-equiripple one- and two-dimensional linear-phase FIR digital filters design approach is proposed based on a novel neural network optimization technique. Its goal is to minimize the weighted square-error function in the frequency domain. The design solution is presented as a parallel algorithm to approximate the desired frequency response specification, and the weight coefficients are updated according to the error function. Thus, the proposed approximation method can avoid the overshoot phenomenon which may happen near the pass-band and stop-band edges of the designed filter, and may make a fast calculation of the filter’s coefficients possible. Several optimal design examples are given to illustrate the effectiveness of the proposed approach.

A neural network based interpolation scheme using the local spatial properties of the source image for image enlargement is proposed. The local spatial properties that are used for neural network training include the neighbor pixels gray values, the average value and the gray value variations between neighbor pixels in the selected region. Gaussian radial basis function neural network is used for image local spatial properties pattern learning and regression estimation for image interpolation. The trained neural network is used to estimate the gray values of unknown pixels using the known neighbor pixels and local spatial properties information. Some interpolation experiments demonstrate that the proposed approach is superior to linear, cubic and other neural network and support vector machines based interpolation approaches.

Image segmentation is a classical problem in the area of image processing, multimedia, medical image, and so on. Although there exist a lot of approaches to perform image segmentation, few of them study the image segmentation by the cluster ensemble approach. In this paper, we propose a new algorithm called the cluster ensemble algorithm (CEA) for image segmentation. Specifically, CEA first obtains two set of segmented regions which are partitioned by EM according to the color feature and the texture feature respectively. Then, it integrates these regions to segmented regions based on the similarity measure and the fuzzy membership function. Finally, CEA performs the denoise algorithm on the segmented regions to remove the noise. The experiments show that CEA works well during the process of image segmentation.

This paper presents a Decision-based Hybrid Image Watermarking (DHIW) technique, based on the Human Visible System (HVS) and an Artificial Neural Network (ANN), for image copyright protection in wavelet domain. In [1], an image watermarking technique, called the IWNN technique, utilizes an ANN to extract watermarks without original images. However, the IWNN technique performs poorly for highly complicated image textures because the generalization capability of neural networks is powerfully effective in dealing with smooth image textures. Therefore, the PAIW method is proposed to enhance the IWNN technique, which uses the spatial information associated with wavelet-transformed images. The DHIW technique takes advantages of these two techniques by using a decision preprocessor. Experimental results prove that the DHIW technique remarkably outperforms other existing schemes.

A new algorithm for magnet resonance (MR) image registration is proposed based on a modified Pulse-Coupled Neural Networks (PCNN’s). The transformed image and reference image are applied as inputs to two modified networks with the same parameters respectively. Taking advantage of translation, rotation, and distortion invariant characteristics of PCNN’s, fired neuron groups of the two networks are acquired correspondingly, then the barycenters of those groups are extracted as characteristic points to attain the registration parameters. Experiment results showed that the proposed algorithm for MR image registration is fast and effective.

Image magnification is among the basic image processing operations. The most commonly used techniques for image magnification are based on interpolation method. However, the magnified images produced by the techniques, such as nearest neighbor, bilinear and cubic method, often appear a variety of undesirable image artifacts such as ’blocking’ and ’blurring’ into the several processing for image magnification. In this paper, we propose image magnification method by properties of human visual system which reduce information during transforming from receptors to ganglion cells in retina and magnify information at visual cortex. Our method uses the whole image to exactly detect the edge information of the image and then emphasizes edge information. Experiment results show that the proposed method solves the drawbacks of the image magnification, such as blocking and blurring, and has a higher PSNR and Correlation than the traditional methods.