Catálogo de publicaciones - libros

The design of an efficient reusable IP based Extended JPEG encoder is presented in this paper. This encoder uses user-defined quantization and Huffman tables that can be reconfigured at run-time. It has a modularized and pipelined architecture with distributed control for each block. A simple interface makes integration of the modules in various systems simple and straightforward. The design when targeted on FPGA operated at speed of up to 90MHz and when mapped on 0.25 μ m CMOS process the design can operate at speeds over 450MHz, which is faster than any of the similar JPEG encoder designs reported.

In network or ubiquitous environments, there are difficulties in performing large vocabulary speech recognition by a small device due to its limited power. Therefore, an approach, so-called distributed speech recognition (DSR), that distributes the processing modules of automatic speech recognition into a device and a server has been attractive. Of all processing modules of DSR, quantization of feature parameters plays a main role in terms of the transmission bandwidth and the recognition performance. In this paper, we propose an efficient quantizer of feature parameters by incorporating the correlation between successive analysis frames of speech. The proposed quantizer is based on the predictive multi-stage vector quantization scheme and designed with different bit rates by trading off with the performance of speech recognition. It is shown from speech recognition experiments that the DSR system employing the proposed quantization method can reduce a bit rate by 20% with a comparable recognition performance to the ETSI DSR standard.

New quantization method based on magnitude-sign split scheme for bandwidth scalable wideband speech codec is proposed. In the high-band codec, the signal is band-pass filtered and each band is transformed independently into DCT domain. The DCT coefficients are split into magnitude and sign, and each is quantized separately based on its unique characteristics. In addition, the quantized gain parameter in the low-band codec is utilized in the high-band codec for an enhanced performance. The 19.8kbps bandwidth scalable wideband codec consisting of G.729E for low-band and the proposed codec for high-band is developed, and it is confirmed that the proposed codec has better subjective performance than 24kbps G.722.1.

In this paper, a high performance asynchronous on-chip bus designed in a Globally Asynchronous Locally Synchronous (GALS) style is proposed. The asynchronous on-chip bus is capable of handling multiple outstanding transactions and in-order completion to achieve a high performance, which is implemented with distributed and modularized control unit in a layered interface. The architecture of asynchronous on-chip bus is discussed and implemented for simulations. Simulation results show that throughput of the proposed asynchronous on-chip bus with multiple outstanding transactions and in-order transaction completion is increased by 31.3%, while power consumption overhead is only 6.76%, as compared to an asynchronous on-chip bus with a single outstanding transaction.

Block-based motion estimation can be regarded as a function minimization problem in a finite two-dimensional space. Therefore, fast block-based motion estimation can be achieved by using an efficient function minimization algorithm instead of using a predefined search pattern, such as the diamond search. The downhill simplex search algorithm is an efficient derivative-free function minimization algorithm. In this paper, we propose several enhanced schemes to improve the efficiency of applying the downhill simplex search algorithm to motion estimation. The proposed enhanced schemes include a new initialization process, a special rounding method, and an early-stop error function evaluation procedure. Experimental results on several benchmarking videos show superior performance of the proposed algorithm over some existing fast block matching methods.

In this paper, we propose a camera motion detection method that can identify pan, tilt and zoom in a video sequence. The proposed method exploits motion features based on the motion cooccurrence matrix, which is able to provide dominant motion characteristics between two images such as the size of the homogeneous motion area and the direction of motion. We show that motion cooccurrence matrices are quite different for different types of motion and can be used to effectively identify simple camera motion such as pan, tilt and zoom in video sequences. Our method does not rely on the parametric motion model and can be used to qualitatively detect camera motion. Performance of the proposed method is evaluated by experiments for a set of test sequence.

We compare, both objectively and subjectively, the performance of various advanced motion compensation methods, including overlapped block motion compensation (OBMC) and control grid interpolation (CGI), in a 3-D wavelet based coding system. Our results indicate that an OBMC sequence usually has a higher PSNR than those of the other methods, while a CGI sequence usually has the least number of blocking artifacts. We combine these two methods in a hybrid system in order to achieve better visual quality and maintain satisfactory coding efficiency simultaneously. The objective and subjective results indicate that the proposed hybrid method removes more than 50% of blocking artifacts of an OBMC sequence, while simultaneously maintaining a high PSNR performance.

We present a method of building a panoramic image from adjacent images. The panoramic image is constructed using several images taken by adjacent cameras or parsed from a video, and used for photogrammetry or many computer graphic applications. The perspective transformation, which is estimated from the appropriate corresponding pairs of images, can be used to construct the panoramic image without unwarranted distortion. We used the corner points for the corresponding features, and morphological structures were utilized for fast and robust corner detection. We used the criterion of the corner strength, which guarantees the robust detection of the corner in most situations. For the transformation, 8 parameters were estimated from perspective equations, and bilinear color blending was used to construct a seamless panoramic video. The experiments showed that the proposed method yields fast results with good quality under various conditions.

This paper presents a new 3D line segment extraction method, which can be used in generating 3D rooftop model. The core of our method is that 3D line segment is extracted by using line fitting of elevation data on 2D line coordinates of ortho-image. In order to use elevation in line fitting, the elevation itself should be reliable. To measure the reliability of elevation, in this paper, we employ the concept of self-consistency. We test the effectiveness of the proposed method with a quantitative accuracy analysis using synthetic images generated from Avenches data set of Ascona aerial images. Experimental results indicate that our method generates 3D line segments almost 10 times more accurate than raw elevations obtained by area-based method. Also, our proposed method shows much improved accuracy over the cooperative hybrid stereo method. Using a simple 3D line grouping scheme, 3D line segments are shown to generate a precise 3D building model effectively.

Automatic liver segmentation from abdominal computed tomography (CT) images is one of the most important steps for computer-aided diagnosis (CAD) for liver CT. However, the liver must be separated manually or semi-automatically since surface features of the liver and partial-volume effects make automatic discrimination from other adjacent organs or tissues very difficult. In this paper, we present an unsupervised liver segmentation algorithm with three steps. In the preprocessing, we simplify the input CT image by estimating the liver position using a prior knowledge about the location of the liver and by performing multilevel threshold on the estimated liver position. The proposed scheme utilizes the multiscale morphological filter recursively with region-labeling and clustering to detect the search range for deformable contouring. Most of the liver contours are positioned within the search range. In order to perform an accurate segmentation, we produce the gradient-label map, which represents the gradient magnitude in the search range. The proposed algorithm performed deformable contouring on the gradient-label map by using regular patterns of the liver boundary. Experimental results are comparable to those of manual tracing by radiological doctors and shown to be efficient.