Paul Bao

Canny edge detection enhancement by scale multiplication

The technique of scale multiplication is analyzed in the framework of Canny edge detection. A scale multiplication function is defined as the product of the responses of the detection filter at two scales. Edge maps are constructed as the local maxima by thresholding the scale multiplication results. The detection and localization criteria of the scale multiplication are derived. At a small loss in the detection criterion, the localization criterion can be much improved by scale multiplication. The product of the two criteria for scale multiplication is greater than that for a single scale, which leads to better edge detection performance. Experimental results are presented.

Edge detection by scale multiplication in wavelet domain

This paper proposes a wavelet based edge detection scheme by scale multiplication. The dyadic wavelet transforms at two adjacent scales are multiplied as a product function to magnify the edge structures and suppress the noise. Unlike many multiscale techniques that first form the edge maps at several scales and then synthesize them together, we determined the edges as the local maxima directly in the scale product after an efficient thrsholding. It is shown that the scale multiplication achieves better results than either of the two scales, especially on the localization performance. The dislocation of neighboring edges is also improved when the width of detection filter is set large to smooth noise. Experiments on natural images are compared with the Laplacian of Gaussian and Canny edge detection algorithms.

The strongest schema learning GA and its application to multilevel thresholding

The multilevel thresholding segmentation methods often outperform the bi-level methods. However, their computational complexity will also grow exponentially as the threshold number increases due to the exhaustive search. Genetic algorithms (GAs) can accelerate the optimization calculation but suffer drawbacks such as slow convergence and easy to trap into local optimum. Extracting from several highest performance strings, a strongest scheme can be obtained. With the low performance strings learning from it with a certain probability, the average-fitness of each generation can increase and the computational time will improve. On the other hand, the learning program can also improve the population diversity. This will enhance the stability of the optimization calculation. Experiment results showed that it was very effective for multilevel thresholding.

Hybrid inter- and intra-wavelet scale image restoration

This paper exploits both the inter- and intra-scale interdependencies that exist in wavelet coefficients to improve image restoration from noise-corrupted data. Using an over-complete wavelet expansion, we group the wavelet coefficients with the same spatial orientation at several scales. We then apply the linear minimum mean squared-error estimation to smooth noise. This scheme exploits the inter-scale correlation information of wavelet coefficients. To exploit the intra-scale dependencies, we calculate the co-variance matrix of each vector locally using a centered square-shaped window. Experiments show that the proposed hybrid scheme significantly outperforms methods exploiting only the intra- or inter-scale dependencies. The performance of noise removal also depends on wavelet filters. In our experiments a biorthogonal wavelet, which best characterizes the image inter-scale dependencies, achieves the best results.

Denoising by spatial correlation thresholding

This paper presents a spatial-correlation thresholding scheme for noise reduction by wavelet transform. Observing that edge structures are of high magnitude across wavelet scales but noise decays rapidly, we multiply two adjacent wavelet scales to form a spatial-correlation function to enhance significant structures and dilute noise. Dissimilar to the traditional thresholding schemes that apply threshold to the wavelet coefficients, the proposed scheme applies threshold directly to the scale correlation. A robust threshold is presented and experiments show that the proposed scheme outperforms the traditional thresholding methods.

L/sub /spl infin//-constrained high-fidelity image compression via adaptive context modeling

We study high-fidelity image compression with a given tight bound on the maximum error magnitude. We propose some practical adaptive context modeling techniques to correct prediction biases caused by quantizing prediction residues, a problem common to the current DPCM like predictive nearly-lossless image coders. By incorporating the proposed techniques into the nearly-lossless version of CALIC, we were able to increase its PSNR by 1 dB or more and/or reduce its bit rate by ten per cent or more. More encouragingly, at bit rates around 1.25 bpp our method obtained competitive PSNR results against the best wavelet coders, while obtaining much smaller maximum error magnitude.

Audio-to-image wavelet transform based audio steganography

In this paper, we propose an audio steganographic scheme based on wavelet audio-to-image transform. The scheme converts the audio steganographic issue into well-explored image steganographic one. In the scheme, the host audio signal is transformed into image, the covert data are embedded in the image by an image steganographic scheme and finally, the image is transformed back into audio signal. The performance of the proposed scheme under MP3 compression is shown in the experimental results

MP3-resistant music steganography based on dynamic range transform

In this paper, we present a novel audio steganography scheme for embedding high-capacity covert data in a music carrier, where the carrier is first filtered into a sub-signal, insensitive to the human auditory system, using a dynamic range filter, then the sub-signal is downsampled and transformed to a 2D arrangement (image), and finally the image is transformed to a wavelet domain singular value decomposition (SVD) on which a quantization-index-modulation process is applied for embedding the covert data. The proposed scheme, due to its indirect embedding on the singular values of the dynamic range filtered data in the visual domain, possesses a high and flexible embedding capacity and a superb MP3 robustness while retaining an excellent inaudibility. The MP3 robustness and the imperceptibility are further enhanced by adaptively modeling the quantization parameters, based on the statistics within subbands. The embedded data extraction is also secure and oblivious.

A wavelet-based edge detection method by scale multiplication

A wavelet-based multiscale edge detection scheme is presented in this paper. By multiplying the wavelet coefficients at two adjacent scales to magnify significant structures and suppress noise, we determined edges as the local maxima directly in the scale product after an efficient thresholding, instead of first forming the edge maps at several scales and then synthesizing them together, as employed in many multiscale techniques. It is shown that the scale multiplication achieves better results than either of the two scales, especially on the localization performance. Experiments on natural images are compared with the Laplacian of Gaussian and Canny edge detection algorithms.

Enhancing the estimation of plant Jacobian for adaptive neural inverse control

Abstract To implement the specialized learning of the inverse dynamic neuro-controller for controlling nonlinear plants with noise, it is strongly desirable that an on-line trained neural plant emluator may provide a reasonably good estimation of the plant Jacobian under noise environments. This paper presents an approach for enhancing the estimation of the plant Jacobian which is on-line used in direct adaptive neural inverse control schemes. The estimated teaching signals are obtained by using the input–output data available at each time step, and then they are used to train the neural plant emluator by a new cost function introduced in this work for training the plant emluator. Convergence theorem for the adaptive back-propagation algorithm and stability of the closed-loop control system are established by using the Lyapunov theory. Simulations are conducted for demonstrating the effectiveness of the proposed strategy.