Wing-Kuen Ling

Local information-based fast approximate spectral clustering

Spectral clustering has become one of the most popular clustering approaches in recent years. However, its high computational complexity prevents its application to large-scale datasets. To address this complexity, approximate spectral clustering methods have been proposed. In these methods, computational costs are reduced by using approximation techniques, such as the Nystrom method, or by constructing a smaller representative dataset on which spectral clustering is performed. However, the computational efficiency of these approximation methods is achieved at the cost of performance degradation. In this paper, we propose an efficient approximate spectral clustering method in which clustering performance is improved by utilizing local information among the data, while the scalability to the large-scale datasets is retained. Specifically, we improve the approximate spectral clustering method in two aspects. First, a sparse affinity graph is adopted to improve the performance of spectral clustering on the small representative dataset. Second, local interpolation is utilized to improve the extension of the clustering result. Experiments are conducted on several real-world datasets, showing that the proposed method is efficient and outperforms the state-of-the-art approximate spectral clustering algorithms.

Linear vs. Nonlinear Extreme Learning Machine for Spectral-Spatial Classification of Hyperspectral Images

As a new machine learning approach, the extreme learning machine (ELM) has received much attention due to its good performance. However, when directly applied to hyperspectral image (HSI) classification, the recognition rate is low. This is because ELM does not use spatial information, which is very important for HSI classification. In view of this, this paper proposes a new framework for the spectral-spatial classification of HSI by combining ELM with loopy belief propagation (LBP). The original ELM is linear, and the nonlinear ELMs (or Kernel ELMs) are an improvement of linear ELM (LELM). However, based on lots of experiments and much analysis, it is found that the LELM is a better choice than nonlinear ELM for the spectral-spatial classification of HSI. Furthermore, we exploit the marginal probability distribution that uses the whole information in the HSI and learns such a distribution using the LBP. The proposed method not only maintains the fast speed of ELM, but also greatly improves the accuracy of classification. The experimental results in the well-known HSI data sets, Indian Pines, and Pavia University, demonstrate the good performance of the proposed method.

Extreme Sparse Multinomial Logistic Regression: A Fast and Robust Framework for Hyperspectral Image Classification

Although sparse multinomial logistic regression (SMLR) has provided a useful tool for sparse classification, it suffers from inefficacy in dealing with high dimensional features and manually set initial regressor values. This has significantly constrained its applications for hyperspectral image (HSI) classification. In order to tackle these two drawbacks, an extreme sparse multinomial logistic regression (ESMLR) is proposed for effective classification of HSI. First, the HSI dataset is projected to a new feature space with randomly generated weight and bias. Second, an optimization model is established by the Lagrange multiplier method and the dual principle to automatically determine a good initial regressor for SMLR via minimizing the training error and the regressor value. Furthermore, the extended multi-attribute profiles (EMAPs) are utilized for extracting both the spectral and spatial features. A combinational linear multiple features learning (MFL) method is proposed to further enhance the features extracted by ESMLR and EMAPs. Finally, the logistic regression via the variable splitting and the augmented Lagrangian (LORSAL) is adopted in the proposed framework for reducing the computational time. Experiments are conducted on two well-known HSI datasets, namely the Indian Pines dataset and the Pavia University dataset, which have shown the fast and robust performance of the proposed ESMLR framework.

Design of convolution neural network with frequency selectivity for wearable camera embed glasses based image recognition systems via nonconvex functional inequality constrained sparse optimization approach

As there is a rapid development of the wearable camera embed glasses in this decade and these wearable camera embed glasses are portable for the consumer uses, many image recognition systems are developed based on these wearable camera embed glasses. To perform the image recognition, a deep learning based convolution neural network is employed. Instead of using the conventional back propagation approach for training the weight matrices in the convolution layer of the convolution neural network, this paper proposes an optimization approach for the design of these weight matrices. In particular, the error energy between the filtered input vectors and the desirable output vectors of the convolution layer as well as the Lp norm of the weight matrices are minimized subject to the frequency selectivity specifications imposed on these weight matrices. This design problem is actually a nonconvex functional inequality constrained sparse problem. Our recently developed sparse optimization method and nonconvex functional inequality constrained optimization method are applied for finding the solution of the optimization problem.

Optimal design of both rectified layer and pooling layer of convolutional neural network for noninvasive blood glucose estimation system

This paper proposes the optimal designs of both the rectified layer and the pooling layer of the convolutional neural network for a non-invasive blood glucose estimation system. The activation function of the neuron in the rectified layer is modelled by a high dimensional Gaussian function. The optimal design of the rectified layer becomes the optimal design of the parameters in the high dimensional Gaussian function. On the other hand, the pooling layer of the convolutional neural network is to represent a certain number of the outputs of the rectified layer by a value. In this paper, this representation value is defined as the Lp norm of a certain number of the outputs of the rectified layer, and the value of p is found via finding the solution of a smooth optimization problem. By finding the solutions of these optimization problems, the designed convolutional neural network is used in a non-invasive system for estimating the blood glucose concentration.

Grouping and Selecting Singular Spectral Analysis Components for Denoising Based on Empirical Mode Decomposition via Integer Quadratic Programming

This study proposes an integer quadratic programming method for grouping and selecting the singular spectral analysis components based on the empirical mode decomposition for performing the denoising. Here, the total number of the grouped singular spectral analysis components is equal to the total number of the intrinsic mode functions. The singular spectral analysis components are assigned to the group indexed by the corresponding intrinsic mode function where the two norm error between the corresponding intrinsic mode function and the sum of the grouped singular spectral analysis components is minimum. Actually, this assignment of the singular spectral analysis components to a particular group is an integer quadratic programming problem. However, the required computational power for finding the solution of the integer quadratic programming problem is high. On the other hand, by representing the integer quadratic programming problem as an integer linear programming problem and employing an existing numerical optimisation computer aided design tool for finding the solution of the integer linear programming problem, the solution can be found efficiently. Computer numerical simulation results are presented.

Preventing potential fires and hazardous situations in consumer products

Every year many consumer products such as TV sets and white goods are reported to have caught fire. In 2012, the Tokyo fire department reported that about 17% of household fires originate from consumer products. Although these figures are much better than in the past, they are still unsatisfactory. Why do consumer products become hazardous despite passing stringent testing? One of the reasons is that consumer products are generally produced in large quantities. Design evaluation and safety testing, however is only done on a very small sample, which does not represent the statistical population of the product. This paper reviews the issue of design robustness, and tolerance issues associated with large quantities such as are present in consumer products. Then it shows how statistical techniques can be adapted to safety engineering to reduce the risk of a hazardous situation occurring.

Big data and cloud based wearable forgery note recognition systems via contourlet transform and support vector machine

Forgery note recognition plays an important role in our daily life as the general public has an opportunity to receive the forgery notes. This paper proposes a big data and cloud based wearable forgery note recognition system. First, a wearable camera is employed for taking the note images. Then, the note images are sent to the cloud systems. There is a big data system in the cloud server. Image features are extracted based on the coutourlet transform. The coutourlet transform is employed for the feature extraction because the contourlet has the multi-resolution property as well as performs the local and directional image decomposition. In particular, the statistical information of the coefficients obtained by performing the contourlet transform is employed as the features. Then, a support vector machine is used for performing the forgery note recognition. Extensive computer numerical simulations are performed. The obtained results show that the proposed method achieves a higher accuracy rate compared to the existing methods.

Empirical relationships between artificial noises and audio performances of wireless industrial audio systems with dithers

This paper presents the empirical relationships between various types of noises and the audio performances of various types of dithering systems based on extensive computer numerical simulations. There are two types of dithering systems. They are the subtractive dithering systems and the nonsubtractive dithering systems. For the subtractive dithering systems, artificial noises are added before and after the quantizer. For the nonsubtractive dithering systems, there is only a single artificial noise added before the quantizer. Also, there are two types of subtractive dithering systems. They are the synchronous subtractive dithering systems and the asynchronous subtractive dithering systems. The artificial noises added before and after the quantizer are the same for the synchronous subtractive dithering systems, while they are different for the asynchronous subtractive dithering systems. In this paper, both the subtractive dithering systems and the nonsubtractive dithering systems as well as both the synchronous subtractive dithering systems and the asynchronous subtractive dithering systems are studied. Besides, the Gaussian distributed noises, the uniform distributed noises and the bilaterial exponential distributed noises as well as the sinusoidal input signals are employed for evaluating both the signal to noise ratios (SNRs) and the tunal suppression ratios (TSRs). Computer numerical simulation results show that the synchronous subtractive dithering systems outperform the nonsubtractive dithering systems. Also, the use of the uniform distributed noises outperforms the use of the other two types of noises. Moreover, the asynchronous subtractive dithering systems achieve constant TSRs independent of both the type of the noises and the SNR levels.

Stabilization of single bit high order interpolative sigma delta modulators for analog-to-digital conversion in wireless mobile handset based electromyogram acquisition system

Wireless mobile handsets are widely used for the electromyogram acquisitions because of their portability property. However, as the electromyograms are with low amplitudes and they are very sensitive to the noises, a good analog-to-digital conversion system plays a very important role for the processing of the electromyograms. Among all the analog-to-digital converters, the sigma delta modulator is the most common analog-to-digital convertor employed in the mobile handsets. This is because the oversampling mechanism for bandlimited signals can be implemented using the existing hardware. However, the sigma delta modulator may suffer from the stability issue. This paper considers the stabilization of a single bit high order interpolative sigma delta modulator via flipping some values of the quantizer output. The values of the quantizer output are determined in such that certain frequency contents of the input signal are cancelled by those of the quantizer output. A stability condition for the proposed control strategy is derived. Since only some frequency detectors and a simple inverter are required for the implementation of the proposed control strategy, the implementation cost is low.