Y. K. Chan

Morphological representation of DCT coefficients for image compression

Recent success in discrete cosine transform (DCT) image coding is mainly attributed to recognition of the importance of data organization and representation. Currently, there are several competitive DCT-based coders such as DCT-based embedded image coding (EZDCT) (see Xiong et al., Z., 1996) and significance tree quantization (STQ) (see Davis, G. and Chawla, S., 1997). In the wavelet context, morphological representation of wavelet data has achieved the best compression performance. The representatives are morphological representation of wavelet data (MRWD) (see Servetto, S. et al., 1999) and significance-linked connected component analysis (see Chai, B.-B. et al., 1999). We show that, by proper reorganization of its coefficients, the block-based DCT can have similar characteristics, such as energy compaction, cross-subband similarity, decay of magnitude across subband, etc., to the wavelet transform. These characteristics can widen DCT applications relevant to image compression, image retrieval, pattern recognition, etc. We then present an image coder utilizing these characteristics by morphological representation of DCT coefficients (MRDCT). The experiments show that MRDCT is among the state-of-the-art DCT-based image coders reported in the literature. For example, for the Lena image at 0.25 bpp, MRDCT outperforms JPEG, STQ and EZDCT in peak signal-to-noise ratio by 1.0, 1.0, and 0.3 dB, respectively.

Image Fusion Based Visible Watermarking Using Dual-Tree Complex Wavelet Transform

Digital watermarking has been researched extensively due to its potential use for data security and copyright protection. Much of the literature has focused on developing invisible watermarking algorithms. However, not much has been done on visible watermarking. A visible watermark is apparently needed for copyright notification. This work proposes a new framework of visible watermarking based on image fusion, a common technique used in combining images acquired from different modalities. To better protect the host features and increase the robustness of the watermark, the dual-tree complex wavelet transform (DT-CWT) is used. A new classification strategy is proposed to classify complex wavelet coefficients into 6 classes with different perceptual significance. The embedding decision can be made based on the classification information. Small watermark coefficients are prohibited from embedding. In the host edges, the insertion of watermark energy is controlled by using the inversely proportional embedding scheme to better preserve the sensitive region, while in other regions, the embedding strength becomes stronger as texture activity increases. This work also addresses the problem of low-pass subband watermark embedding, which is a special issue of visible watermarking. Experimental results show that the proposed algorithm yields significantly superior image quality than the current DCT-based method.

The design and application of DWT-domain optimum decoders

Informed embedding is the practice of tailoring each watermarking pattern according to the cover Work in which it is to be embedded. The objective is to attain an optimal trade-off between estimates of perceptual fidelity and robustness. To date, our own studies of informed embedding have been limited to watermarks with very small data payloads. Our purpose in the present paper is to develop a method of informed embedding that is suitable for watermarks with large data payloads. The method we develop employs an estimate of robustness based on the amount of white noise that can be added before a message error becomes likely. We present an iterative, Monte-Carlo algorithm that tries to ensure watermarks are embedded with a specified value of this robustness estimate. This algorithm is tested in an image watermarking system, and is found to successfully embed robust, 129-bit watermarks in 368 × 240 images.

HMM adaptation techniques in training framework

This paper presents an adaptation approach based on the Baum-Welch algorithm method. This method applies the same framework as is are used for training speech recognizers with abundant training data. The Baum-Welch adaptation method is adapted to all the parameters of the hidden Markov models (HMM) with adaptation data. If a large amount of adaptation data is available, these methods could gradually approximate the speaker-dependent ones. The approach is evaluated through the phoneme recognition task on the TIMIT corpus. On the speaker adaptation experiments, up to 91.48% recognition rate is achieved.

Extending DACLIC for near-lossless compression with postprocessing of greyscale images

Summary form only given. A lossless/near lossless coding scheme, DACLIC, is presented. The proposed scheme attempts to remove redundancy in a given image in the spatial domain. The redundancy removal is achieved by block direction prediction and context-based error modeling. The block direction operation in DACLIC first partitions an image into blocks. Pixels within each incoming block are analyzed resulting in a best directional prediction for that block. The best direction is chosen from a given set that results in the minimum prediction error. Removal of redundancy by the block direction technique is not possible for removing all possible redundancy in a given image. Another decorrelation part of the DACLIC scheme is the context-based error modeling which exploits context-dependent DPCM error structures. The DACLIC scheme is primarily used as a lossless image compression technique. However, the scheme can be easily extended to near-lossless compression applications by introducing a small quantization loss. This small quantization loss is restricted to an absolute error not exceeding a prescribed value n for all pixels in a given image. Application of block direction and context modeling reduces a given image into residuals. This residual typically has a lower entropy than the given image. A quadtree Rice coder (QRC) is proposed as an entropy coder of DACLIC. An arithmetic coder is also given as an option. The QRC operates on residual blocks with low computing complexity that compares favorably with the residual coding method used by LOCO-I as the proposed QRC is two-dimensional in nature. For near-lossless compression with a larger value of n, banding artifacts are visible in the decoded image. In the DACLIC system, a postprocessing technique is proposed to remove the banding artifacts.

An Alternative Superscalar Architecture with Integer Execution Units Only

Superscalar architecture resulting in aggressive performance is a proven architecture for general purpose computation. The negative side effect of aggressive performance is the need for higher number of register read/write ports to supply operands to multiple execution units; the need to resolve false data dependence and true data dependence; the need to dispatch operand ready instructions to execution units and finally retire out of order executed instructions to program order. A processor architecture is proposed in here to address at least some of the above negative side effects. This processor architecture is call LITERAL QUEUE ARCHITECTURE(LQA). In here, LITERAL has the meaning of immediate data. In LQA, opcode and operands in an instruction are treated as a self contained structured element and forms the necessary and sufficient condition for instruction execution. Sequence of instructions embedded with LITERALS are treated as elements in a QUEUE. The elements are executed with respect to time and rotated out of the QUEUE while new elements are rotated into the QUEUE.

LLEC: An Image Coder with Low-Complexity and Low-Memory Requirement

A Low-complexity and Low-memory Entropy Coder (LLEC) for image compression is proposed in this paper. The two key elements in LLEC are zerotree coding and Golomb-Rice codes. Zerotree coding exploits the zerotree structure of transformed coefficients for higher compression efficiency. Golomb-Rice codes are used to code the remaining coefficients in a VLC/VLI manner for low complexity and low memory. The experimental results show that the compression efficiency of DCT- and DWT-based LLEC outperforms baseline JPEG and EZW at the given bit rates, respectively. When compared with SPIHT, LLEC is inferior by 0.3 dB on average for the tested images but superior in terms of computational complexity and memory requirement. In addition, LLEC has other desirable features such as parallel processing support, ROI (Region Of Interest) coding and as a universal entropy coder for DCT and DWT.

FACOLA-face coder based on location and attention

Summary form only given. FACOLA (face coder based on location and attention) is proposed for potential applications such as compression of face pictures used in IC and ID cards. The face locator locates a face in an image using template matching and eigenface techniques. The attention detector detects high attention and low attention in the located face according to their different frequency characteristics. For high attention, usually high quality or lossless coding is required. The DCT is not suitable for such a case because it will not lead to an efficient compaction of the image energy and variable length coding (VLC) cannot be done efficiently. So a DPCM coder with three directional predictors is presented instead of the DCT. The prediction difference is quantized and the entropy coded DPCM coder supports lossless and lossy compression. The DCT coder is adopted for low attention and non-face area compression using different quantization factors.

Environment for Three Dimensional Graphics User Interface Development

Graphics User Interface(GUI) serves as a link between the end user of an application and the application itself. Run time GUI is characterized by its look-and-feel and interaction style. Characteristics of a run time GUI system depends on the particular GUI building tool used and its underlying architecture. Current implementation of GUI are in two-dimensions(2D) and the trend is to extend GUI into three-dimensions(3D) in order to provide a natural interface for applications in 3D. It is expected that 3D specific GUI builder will become available when methodology for automating 3D GUI building process for a large class of 3D application is better understood. Until then, it is proposed that a 3D development and debugging environment should be available for building such interfaces for individual application as required. This paper describes the overall framework and some details that enables the building of 3D GUI.

Discrete utterance recognition based on nonlinear model identification with single layer neural networks

A scheme for speaker independent discrete utterance recognition using single layer neural network (SLNN) based nonlinear auto regression model parameters as the features is presented. A fast training algorithm is developed for the identification of the model parameters. Dynamic programming is used for the pattern matching. This study demonstrates that the SLNN can be used successfully as the short time nonlinear auto regression model of the speech signal and thus acts as a feature extractor for speech recognition. Ten digits uttered by twelve speakers were used as the database to examine the performance of the SLNN based feature extractor of speech as compared to the standard linear prediction technique.<<ETX>>