Catálogo de publicaciones - libros

Fisherface is a popular subspace algorithm used in face recognition, and is commonly believed superior to another technique, Eigenface, due to its attempt to maximize the separability of training classes. However, the obtained discriminating subspace of the training set may not easily extend to unseen classes (thus poor generalization), as in the case of enrollment of new subjects. In this paper, we reduce the performance variance and improve the generalization of Fisherface by automatically selecting some representative classes for training, using a recently proposed neural network architecture SOM. The experiments on ORL face database validate the proposed method.

In this paper, we propose a new method to improve the image registration accuracy in feedforward neural networks (FNN) based scheme. In the proposed method, Bayesian regularization is applied to improve the generalization capability of the FNN. The features extracted from the image sets by kernel independent component analysis (KICA) technique are input vectors of regularized FNN. The outputs of the neural network are those translation, rotation and scaling parameters with respect to reference and observed image sets. Comparative experiments are performed between FNN with regularization and without regularization under various conditions. The results show that the proposed method is much improved not only at accuracy but also remarkably at robust to noise.

This paper presents a new approach for automatic image color correction, based on statistical learning. The method both parameterizes color independently of illumination and corrects color for changes of illumination. The motivation for using a learning approach is to deal with changes of lighting typical of indoor environments such as home and office. The method is based on learning color invariants using a modified multi-layer perceptron (MLP). The MLP is odd-layered. The middle layer includes two neurons which estimate two color invariants and one input neuron which takes in the luminance desired in output of the MLP. The advantage of the modified MLP over a classical MLP is better performance and the estimation of invariants to illumination. The trained modified MLP can be applied using look-up tables (LUTs), yielding very fast processing. Results illustrate the approach.

The paper proposes a new method for image resolution enhancement from multiple images using the limited recurrent neural network (LRNN) approach, which is a set of collectively operating feed-forward neural networks. In the limited recurrent networks, information about past outputs is fed back through recurrent connections of output units and mixed with the input nodes flowing into the network input as external input nodes. Thus, experience about past search is utilized, which enables LRNN to be capable of both learning and searching the optimal solution for optimization problems in the solution space. Estimates computed from a low-resolution (LR) simulation image sequence and an actual video film sequence show dramatic visual and quantitative improvements over bilinear interpolation, and equivalent performance to that of the frequency domain approach.

With the availability of multi-sensor and multi-frequency image data from operational observation satellites, the fusion of image data has become an important tool in remote sensing image evaluation and segmentation. This paper presents a novel Radius Basis Function (RBF) neural network with some distinctive training strategies, which can integrate multiple information sources efficiently and exploit the potential advantages of each feature. Multi-scale features extracted from remote sensing images are evaluated adaptively and used for segmentation. Experimental results obtained on artificial and real data are both presented which demonstrate the effectiveness of our proposal.

This paper proposes a novel active contour model for image object recognition using neural networks as a dynamic information fusion kernel. It first learns feature fusion strategies from training data by searching for an optimal fusion model at each marching step of the active contour model. A recurrent neural network is then employed to learn the fusion strategy knowledge. The learned knowledge is then applied to guide another linear neural network to fuse the features, which determine the marching procedures of an active contour model for object recognition. We test our model on both artificial and real image data sets and compare the results to those of a standard active model, with promising outcomes.

How to use a polygon with the fewest possible sides to approximate a shape boundary is an important issue in pattern recognition and image processing. A novel split-and-merge technique(SMT) is proposed. SMT starts with an initial shape boundary segmentation, split and merge are then alternately done against the shape boundary. The procedure is halted when the pre-specified iteration number is achieved. For increasing stability of SMT and improving its robustness to the initial segmentation, a ranking-selection scheme is utilized to choose the splitting and merging points. The experimental results show its superiority.

A fast and simple solution was suggested to reduce the inter-pixels correlations in natural images, of which the power spectra roughly fell off with the increasing spatial frequency according to a power law; but the 1/ exponent, , was different from image to image. The essential of the proposed method was to flatten the decreasing power spectrum of each image by using an adaptive low-pass and whitening filter. The act of low-pass filtering was just to reduce the effects of noise usually took place in the high frequencies. The act of whitening filtering was a special processing, which was to attenuate the low frequencies and boost the high frequencies so as to yield a roughly flat power spectrum across all spatial frequencies. The suggested method was computationally more economical than the geometric covariance matrix based PCA method. Meanwhile, the performance degradations accompanied with the computational economy improvement were fairly insignificant.

Concerning the control of plasma column evolution in ITER machine, the reconstruction of the plasma shape in the vacuum vessel represents an important step. In this work, starting from magnetic measurements, a soft computing approach to estimate the distances of the plasma boundary from the first wall of the vacuum vessel is carried out by means of Neural Networks (NNs). In particular, Multi-Layer Perceptron (MLP) nets have been exploited for the purpose. Finally, to verify the robustness of the proposed approach, any different database and number of input parameters has been used.

Functional Magnetic Resonance Imaging (fMRI) requires ultra-fast imaging in order to capture the on-going spatio-temporal dynamics of the cognitive task. We make use of correlations in both -space and time, and thereby reconstruct the time series by acquiring only a fraction of the data, using an improved form of the well-known dynamic imaging technique - BLAST (Broad-use Linear Acquisition Speed-up Technique). - BLAST (-B) works by unwrapping the aliased Fourier conjugate space of - ( - space). The unwrapping process makes use of an estimate of the true - space, obtained by acquiring a blurred unaliased version. In this paper, we propose two changes to the existing algorithm. Firstly, we improve the map estimate using generalized series reconstruction. The second change is to incorporate phase constraints from the training map. The proposed technique is compared with existing -B on visual stimulation fMRI data obtained on 5 volunteers. Results show that the proposed changes lead to gain in temporal resolution by as much as a factor of 6. Performance evaluation is carried out by comparing activation maps obtained using reconstructed images, against that obtained from the true images. We observe upto 10dB improvement in PSNR of activation maps. Besides, RMSE reduction on fMRI images, of about 10% averaged over the entire time series, with a peak improvement of 35% compared to the existing -B, averaged over 5 data sets, is also observed.