Catálogo de publicaciones - libros

We describe the architecture and implementation of bilateral background subtraction for real-time stereo vision system. Pre-smoothing a signal and noise removal may help to improve the performance for many signal-processing algorithms such as compression, detection, enhancement, recognition, and more [2]. Bilateral filtering proposed by C. Tomasi and R. Manduchi can be used as an edge-preserving smoother, removing high-frequency components of an image without blurring its edges [1][3]. Recently, [3] showed enhanced real-time stereo through software implementation of bilateral filtering. In this paper, we show hardware implementation of bilateral background subtraction for real-time stereo and present its architecture as well as required hardware resources. Also, we provide experimental results with real data and present our future works.

This paper presents a new description of image edge in the form of asymptote equation utilizing knowledge of differential geometry, and modifies a popular parlance in the traditional image processing. This paper develops the gradient concept and defines a generalized grads operator. The operator is a novel nonlinear transform, which inherits the strongpoint of noise suppression from the classical Gaussian differential filter and has a better edge extraction function. The experimental results show the proposed edge detection algorithm is powerful and effective.

In this paper, we present a new image compression algorithm which combines Wavelet Support Vector Machines (WSVM) learning with the wavelet transform. Based on the characteristic of wavelet transform, Daubechies 9/7 wavelet has been used to transform the image and the wavelet coefficients are trained with WSVM using translation-invariant wavelet kernels. Compression is achieved by using WSVM learning to approximate wavelet coefficients with the predefined level of accuracy. A minimal number of coefficients (support vectors) are then encoded by an effective entropy coder based on run-length and arithmetic coding. Experimental results show that the proposed method gains better performance than that of existing compression algorithm.

A framework based on manifold learning in reconstructed state space is proposed as feature extraction means for nonlinear signal classification. On one hand, manifolds are of importance in characterizing chaotic attractors. On the other hand, there are a large number of toolkits in the context of manifold learning. These motivate us to apply manifold learning in reconstructed state space as feature extraction means for nonlinear signal classification, which bridges the gap between nonlinear science and manifold learning and enables a new viewpoint to study nonlinear signals. In this study, the nonlinear signal analysis is performed as follows. First, we embed the time series of interest into a high-dimensional space via state space reconstruction. Then, we employ locally linear embedding (LLE) to obtain the local manifold characteristics around every point in the reconstructed state space. Finally, we summarize all the local features into a global representation via principal component analysis (PCA). Two case studies of oceanic and EEG signal classification were carried out with the proposed scheme. As confirmed by the experiments, the proposed methodology is effective for such applications. This paper puts forward not only a feature extraction method but also a new direction in which a large number of toolkits are available for nonlinear signal analysis for the sake of signal classification.

In this paper, a frame rate up-conversion (FRC) algorithm using the motion vector frequency of neighboring blocks to reduce the block artifacts caused by failure of conventional motion estimation (ME) based on block matching algorithm (BMA) is proposed. The proposed method is based on the Spiral Full Search with early termination, which is applied to avoid the local minima on pattern-like image. Also, the motion vector correction that replaces a low frequency motion vector by a high frequency motion vector of neighborhood is used in the proposed method to reduce the block artifacts caused by the failure of conventional ME. In addition, bi-directional motion compensated interpolation (MCI) using the blocking index to reduce the block artifacts in occlusion area is used in the proposed method. Experimental results show good performance of the proposed scheme with significant reduction of the erroneous motion vectors and block artifacts.

In design and realization of analog electronic circuit, we usually use preferred value components, the performance of practical circuits often deviate from the ideal design target due to rounding the calculated component values to preferred ones. The best combination of the preferred value components exists in general, but the searching space of all combinations of preferred-value components is very huge. Clonal Selection Algorithms (CSA) is a widely used approach for handling optimization problems. In this paper, CSA is applied into searching optimal components for 4th order Butterworth filter design. Simulation results demonstrate that the proposed method is much superior to the conventional means. This method also can be applied into other types of filter design.

Polygonal models have grown rapidly in complexity over recent years, yet most conventional simplification algorithms were designed to handle modest size datasets of a few tens of thousands of triangles. We present a parallel simplification method for large polygonal models. Our algorithm will partition the original model firstly, send each portion to a slave processor, simplify them concurrently, and merge them together lastly. We give an efficient method to deal with the problem of partition border and portion merging. With parallel implementation, the algorithm can handle extremely large data set, and speed up the execution time. Experiment shows that our algorithm can produce approximations of high quality.

According to Cachin’s steganography security criterion, if the statistical distributions of cover and stego images are identical, the hidden message is assumed undetectable. However, any steganographic method will surely cause some statistical distortions, which gives steganalyst a hint. This paper presents a steganalysis method for JPEG images based on Cachin criterion. It estimates the cover image from the stego one by using a small-scale geometrical transform, and then detects the statistical distortions between the cover and stego images based on some features, which are sensitive to the steganographic modifications. Then a classifier is trained on these features. Three different modern steganographic schemes are tested. Experimental results show that the proposed steganalysis scheme has better performance compared to the current steganalysis methods.

Filtering is a preliminary process in many medical image processing applications. It is aiming at reducing noise in images, and any post-processing tasks may benefit from the reduction of noise. The major two noises in computed radiography (CR) images are Gaussian white noise and Poisson noise. By considering both the characteristics of CR images and the statistical features of wavelet transformed coefficients, an efficient spatial adaptive filtering algorithm, which is based on statistical model of local dependency of CR image wavelet coefficients and the approximate minimum mean squared error (MMSE) estimation, is proposed to decrease the Gaussian white noise in computed images. The process is computational cost saving, and the denoising experiments show the algorithm outperforms other approaches in peak-signal-to-noise ratio (PSNR).

A new image coding method based on discrete directional wavelet transform (S-WT) and quadtree decomposition is proposed here. The S-WT is a kind of transform proposed in [1], which is based on lattice theory, and with the difference with the standard wavelet transform is that the former allows more transform directions. Because the directional property in a small region is more regular than in a big block generally, in order to sufficient make use of the multidirectionality and directional vanishing moment(DVM) of S-WT, the input image is divided into many small regions by means of the popular quadtree segmentation, and the splitting criterion is on the rate-distortion sense. After the optimal quadtree is obtained, a resource bit allocation algorithm is fast implemented utilizing the model proposed in [15]. Experiment results indicate that our algorithms perform better compared to some state-of-the-art image coders.