SAS has become an increasingly important public-health problem in recent years. It can adversely affect neurocognitive, cardiovascular, respiratory diseases and can also cause behavior disorder. Moreover, up to 90% of these cases are obstructive sleep apnea (OSA). Therefore, the study of how to diagnose, detect and treat OSA is becoming a significant issue, both academically and medically. Polysomnography can monitor the OSA with relatively fewer invasive techniques. However, polysomnography-based sleep studies are expensive and time-consuming because they require overnight evaluation in sleep laboratories with dedicated systems and attending personnel. Therefore, to improve such inconveniences, one needs to develop a simplified method to diagnose the OSA, so that the OSA can be detected with less time and reduced financial costs. Since currently there seems to be no OSA detection technique available in Taiwan, the goal of this work is to develop a reliable OSA diagnostic algorithm. In particular, via signal processing, feature extraction and artificial intelligence, this thesis describes an on-line ECG-based OSA diagnostic system. It is hoped that with such a system the OSA can be detected efficiently and accurately.

Application of Image inpainting ranges from object removal, photo restoration, scratch removal, and so on. In this thesis, we will propose a modified multi-scale method and learning-based method using artificial neural networks for image inpainting. Multi-scale inpainting method combines image segmentation, contour estimation, and exemplar-based inpainting. The main goal of image segmentation is to separate image to several homogeneous regions outside the target region. After image segmentation, we use contour estimation to estimate curves inside the target region to partition the whole image into several different regions. Then we fill those different regions inside the target region separately by exemplar-based inpainting method. The exemplar-based technique fills the target region via the texture synthesis and filling order of exemplary patches. Exemplary patches are found near target region and the filling order is determined by isophote and densities of exemplary patches. Learning-based inpainting is a novel technique. This technique combines machine learning and the concept of filling order. We use artificial neural networks to learn the structure and texture surrounding the target region. After training, we fill the target region according to the filling order. From our simulation results, very good results can be obtained for removing large-size objects by using the proposed multi-scale method, and for removing medium-size objects of gray images.

In multiresolution analysis(MRA)by wavelet function Daubechies (db), we decompose the signal in two parts, the low and high-frequency content,respectively. We remove the high-frequency content and reconstruct a new ¡§de-noise¡¨ signal by using inverse wavelet transform. In order to improve the forecasting accuracy of ANN (Artificial Neural Network) model ,we use the concept of tidal constituent phase-lags, and the new ¡§de-noise¡¨ signal was used as the input data set of ANN. Besides, we also use wavelet spectrum, conventional energy spectrum (Fast Fourier Transform, FFT),and harmonic analysis to analyze the character of tidal data . The results show that the concept of tidal constituent phase-lags can improve ANN model of tidal forecasting and supplement, but in the wavelet analysis , the improvement is insignificant .The reason is that the energy of higher frequency noise is very small compared to the energy of the diurnal and the semi- diurnal tidal components. In other word , the ANN model has a certain tolerance of noise effect .

Predicting well inflow performance relationship accurately is very important for production engineers. From these predictions, future plans for handling and improving well performance can be established. One method of predicting well inflow performance is to use artificial neural networks. Vogel's reference curve, which is produced from a series of simulation runs for a reservoir model proposed by Weller, is typically used to predict inflow performance relationship for solution-gas-drive reservoirs. In this study, I reproduced Vogel's work, but instead of producing one curve by conventional regression, I built three neural network models. Two models predict the IPR efficiently with higher overall accuracy than Vogel's reference curve.

The aim of this thesis work is to investigate the algorithm of speech emotion recognition using MATLAB. Firstly, five most commonly used features are selected and extracted from speech signal. After this, statistical values such as mean, variance will be derived from the features. These data along with their related emotion target will be fed to MATLAB neural network tool to train and test to make up the classifier. The overall system provides a reliable performance, classifying correctly more than 82% speech samples after properly training.

An automated signal classifier and a semi-automated signal identifier are designed for collecting the dielectrophoretic signatures of cells flowing through a dielectrophoretic cytometer. In past work, the DEP cytometer signals were manually sorted by going through all recorded signals, which is impractical when analyzing 1000’s of cells per day. In the semi-automated method of collection, signals are automatically identified as events and displayed on the user interface to be accepted or rejected by the user. This approach reduced signal collection time by more than half and produced statistics nearly identical to the manual method. The automated signal classifier based on pattern recognition categorizes detection signals as ‘Accept’ or ‘Reject’. Analyzing large volumes of detection signals is possible in much reduced times and may be approaching real time capability.

An automated signal classifier and a semi-automated signal identifier are designed for collecting the dielectrophoretic signatures of cells flowing through a dielectrophoretic cytometer. In past work, the DEP cytometer signals were manually sorted by going through all recorded signals, which is impractical when analyzing 1000’s of cells per day. In the semi-automated method of collection, signals are automatically identified as events and displayed on the user interface to be accepted or rejected by the user. This approach reduced signal collection time by more than half and produced statistics nearly identical to the manual method. The automated signal classifier based on pattern recognition categorizes detection signals as ‘Accept’ or ‘Reject’. Analyzing large volumes of detection signals is possible in much reduced times and may be approaching real time capability.

Deep neural networks are proven to be highly effective tools in various domains, yet their computational and memory costs restrict them from being widely deployed on portable devices. The recent rapid increase of edge computing devices has led to an active search for techniques to address the above-mentioned limitations of machine learning frameworks. The quantization of artificial neural networks (ANNs), which converts the full-precision synaptic weights into low-bit versions, emerged as one of the solutions. At the same time, spiking neural networks (SNNs) have become an attractive alternative to conventional ANNs due to their temporal information processing capability, energy efficiency, and high biological plausibility. Despite being driven by the same motivation, the simultaneous utilization of both concepts has not been fully studied. Therefore, this thesis work aims to bridge the gap between recent progress in quantized neural networks and SNNs. It presents an extensive study on the performance of the quantization function, represented as a linear combination of sigmoid functions, exploited in low-bit weight quantization in SNNs. The given quantization function demonstrates the state-of-the-art performance on four popular benchmarks, CIFAR10-DVS, DVS128 Gesture, N-Caltech101, and N-MNIST, for binary networks (64.05%, 95.45%, 68.71%, and 99.365 respectively) with small accuracy drops (8.03%, 1.18%, 3.47%, and 0.17% respectively) and up to 32x memory savings, which outperforms the existing methods.

Face recognition is the problem of identifying individuals in images. This thesis evaluates two methods used to determine if pairs of face images belong to the same individual or not. The first method is a combination of principal component analysis and a neural network and the second method is based on state-of-the-art convolutional neural networks. They are trained and evaluated using two different data sets. The first set contains many images with large variations in, for example, illumination and facial expression. The second consists of fewer images with small variations. Principal component analysis allowed the use of smaller networks. The largest network has 1.7 million parameters compared to the 7 million used in the convolutional network. The use of smaller networks lowered the training time and evaluation time significantly. Principal component analysis proved to be well suited for the data set with small variations outperforming the convolutional network which need larger data sets to avoid overfitting. The reduction in data dimensionality, however, led to difficulties classifying the data set with large variations. The generous amount of images in this set allowed the convolutional method to reach higher accuracies than the principal component method.

<p>Through the computing advance and GPU (Graphics Processing Unit) availability for math calculation, the deep learning field becomes more popular and prevalent. Object detection with deep learning, which is the part of image processing, plays an important role in automatic vehicle drive and computer vision. Object detection includes object localization and object classification. Object localization involves that the computer looks through the image and gives the correct coordinates to localize the object. Object classification is that the computer classification targets into different categories. The traditional image object detection pipeline idea is from Fast/Faster R-CNN [32] [58]. The region proposal network generates the contained objects areas and put them into classifier. The first step is the object localization while the second step is the object classification. The time cost for this pipeline function is not efficient. Aiming to address this problem, You Only Look Once (YOLO) [4] network is born. YOLO is the single neural network end-to-end pipeline with the image processing speed being 45 frames per second in real time for network prediction. In this thesis, the convolution neural networks are introduced, including the state of art convolutional neural networks in recently years. YOLO implementation details are illustrated step by step. We adopt the YOLO network for our applications since the YOLO network has the faster convergence rate in training and provides high accuracy and it is the end to end architecture, which makes networks easy to optimize and train. </p>