Robust Deep Age Estimation Method Using Artificially Generated Image Set

Human age estimation is one of the key factors in the field of Human–Robot Interaction/Human–Computer Interaction (HRI/HCI). Owing to the development of deep‐learning technologies, age recognition has recently been attempted. In general, however, deep learning techniques require a large‐scale database, and for age learning with variations, a conventional database is insufficient. For this reason, we propose an age estimation method using artificially generated data. Image data are artificially generated through 3D information, thus solving the problem of shortage of training data, and helping with the training of the deep‐learning technique. Augmentation using 3D has advantages over 2D because it creates new images with more information. We use a deep architecture as a pre‐trained model, and improve the estimation capacity using artificially augmented training images. The deep architecture can outperform traditional estimation methods, and the improved method showed increased reliability. We have achieved state‐of‐the‐art performance using the proposed method in the Morph‐II dataset and have proven that the proposed method can be used effectively using the Adience dataset.     

Neural-Network-Based Sensor Fusion of Optical Emission and Mass Spectroscopy Data for Real-Time Fault Detection in Reactive Ion Etching

To achieve timely and accurate fault detection in reactive ion etching, neural networks (NNs) have been applied for the fusion of data generated by two in-situ sensors: optical emission spectroscopy (OES) and residual gas analysis (RGA). While etching is performed, OES and RGA data are simultaneously collected in real time. Several pre-determined, statistically significant wavelengths (for OES data) and atomic masses (for RGA signals) are monitored. These data are subsequently used for training NN-based time series models of process behavior. Such models, referred to herein as time series NNs (TSNNs), are realized using multilayered perceptron NNs. Results indicate that the TSNNs not only predict process parameters of interest, but also efficiently perform as sensor fusion of the in-situ sensor data.     

A soft computing approach for rainfall retrieval from the TRMM microwave imager

A neural network model for rainfall retrieval over ocean from remotely sensed microwave (MW) brightness temperature (BT) is proposed. BT data are obtained from the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI). The BT values from different channels of TMI over the Pacific Ocean (163/spl deg/ to 177/spl deg/W and 18/spl deg/ to 34/spl deg/S) are the input features. The near-surface rainfall rate from the Precipitation Radar (PR) are considered as a target. The proposed model consists of a neural network with online feature selection (FS) and clustering techniques. A K-means clustering algorithm is applied to cluster the selected features. Different networks have been trained to give an instantaneous rainfall rate with all input features as well as with selected features obtained by applying the FS algorithm. It is found that the hybrid network utilizing FS and clustering techniques performs better. The developed network is also validated with two independent datasets on March 14, 2000 over the Atlantic Ocean having stratiform rain and on March 21, 2000 over the Pacific Ocean having both stratiform and convective rain. In both cases, the hybrid network performs well with correlation coefficient improving to 0.78 and 0.81, respectively, in contrast to 0.70 and 0.75 for the network with all features. The rainfall rate retrieved from the hybrid network is also compared with the TMI surface rain rate, and a correlation of 0.84 and 0.75 is found for the two events. The proposed hybrid model is validated with a Doppler Weather Radar, and correlation of 0.52 is observed.     

基于多目标联合优化的语音增强方法研究

语音增强旨在从受噪声干扰的语音中提取目标语音,目前基于神经网络的语音增强方法在提升语音质量和可懂度方面已被证明是有效的。通过多目标联合优化,利用不同特征之间的互补性,可以提升基于神经网络的语音增强方法的性能。然而,这类多目标学习的语音增强方法在网络优化过程中,通常分别对单个输出目标进行损失函数的计算,多目标之间是并行的,并没有充分利用多目标之间可能存在的关联。为了在网络训练过程中增加输出目标间的关联,本文利用长短时记忆网络构建一种双输出系统框架,设计一种多目标损失函数计算策略用于网络训练。该框架估计出目标语音和噪声,基于此得到估计的带噪语音,然后对这三部分进行联合优化。实验结果表明,所提方法可以提高网络对噪声抑制能力,通过该策略可以获得质量更高,噪声残留更少的增强语音。      

基于SVM模型的单目红外图像深度估计

提出一种通过非线性学习模型来估计单目红外图像深度的算法。该算法首先通过逐步线性回归和独立成分分析(ICA)寻找对于红外图像深度相关性较强的特征,然后以具有核函数的非线性支持向量机(SVM)为模型基础,采用监督学习的方法对红外图像深度特征进行回归分析并训练,在训练过程中通过已知数据回归后的最小均方误差对模型参数进行修正,训练后的模型可对单目红外图像的深度分布进行估计。实验结果证明,利用该模型能较一致地估计单目红外图像的深度信息。     

Nonparametric regression analysis achieved with topographic mapsdeveloped in combination with projection pursuit learning: anapplication to density estimation and adaptive filtering of grey-scaleimages

A novel approach to nonparametric regression analysis using topographic maps is proposed. The maps are trained with the extended maximum entropy learning rule (eMER) in combination with projection pursuit regression (PPR) learning. Rather than a single map, several maps are developed along optimally chosen projection directions in the input space. In this way, the regression performance improves in the case of sparsely sampled input spaces. We explore two applications of the eMER/PPR combination: (1) probability density estimation from pilot estimates and (2) adaptive filtering of grey-scale images. The first case is used as a testbed for comparing different, both classic and neural network-based, regression techniques. The results show that our eMER/PPR combination yields a superior regression performance for small data sets. In the second case, the regression model is trained on a noisy subimage. The model obtained after training reduces the noise content of the full image by more than 20 dB     

Estimating rainfall using machine learning strategies based on weather radar data

The precision of forecasting rainfall is vital owing to current world climate change. As deterministic weather forecasting models are usually time consuming, it becomes challenging to efficiently use this large volume of data in hand. Machine learning methods are already proven to be good replacement for traditional deterministic approaches in weather prediction. This paper presents an approach using recurrent neural networks (RNN) and long short term memory (LSTM) techniques to improve the rainfall forecast performance. This will be compared with the random forest classifier and XGBoost as well. The goal is to predict a set of hourly rainfall levels from sequences of weather radar measurements. Python libraries are utilized to forecast the time series data. The training set comprises of data from first 20 days of every month and the inference set data from the continuing days. This makes sure that both train and inference sets are more or less independent. The idea resides in implementing an end‐to‐end learning framework.     

Neural-style microsystems that learn

The basic operation of biological and electronic (artificial) neural networks (NNs) is examined. Learning by NNs is discussed, covering supervised learning, particularly back-propagation, and unsupervised and reinforcement learning. The use of VLSI implementation to speed learning is considered briefly. Applications of neural-style learning chips to pattern recognition, data compression, optimization, and expert systems is discussed. Problem areas and issues for further research are addressed     

A two-dimensional satellite rainfall error model

A two-dimensional satellite rainfall error model ( SREM2D) is developed for simulating ensembles of satellite rain fields on the basis of "reference" rain fields derived from higher accuracy sensor estimates. With this model we aim at characterizing the multidimensional stochastic error structure of satellite rainfall estimates as a function of scale. The pertinent error dimensions we seek to address are: 1) the joint probability density function for characterizing the spatial structure of the successful delineation of rainy and nonrainy areas; 2) the temporal dynamics of rain estimation bias; and 3) the spatial variability of rain rate estimation error. Ground radar rain fields in the Southern plains of the United States are used as reference to evaluate SREM2D error parameters at 0.25-deg and hourly spatiotemporal resolution for an infrared (IR) rain retrieval algorithm (IR-3B41RT) developed at NASA. Comparison of SREM2D simulated satellite rainfall with actual IR-3B41RT data showed that the error modeling technique can preserve the estimation error characteristics across scales with marginal deviations. The model performance is compared against two simpler, but widely used, approaches of error modeling that do not account for uncertainty in rainy/nonrainy area delineation. It is shown that both of these approaches fare poorly with regards to preserving the error structure across scales. They underestimated the sensor retrieval error standard deviation by more than 100% upon aggregation, which, for SREM2D, was found to be below 30%. SREM2D is modular in design-it can be applied for any satellite rainfall algorithm to consistently characterize its error structure.     

Three-dimensional modeling and visualization of the cochlea on theInternet

3D modeling and visualization of the cochlea using the World Wide Web (WWW) is an effective way of sharing anatomic information for cochlear implantation over the Internet, particularly for morphometry-based research and resident training in otolaryngology and neuroradiology. In this paper, 3D modeling, visualization and animation techniques are integrated in an interactive and platform-independent manner and implemented over the WWW. L.T. Cohen et al.'s (1996) template shape with mean cross-sectional areas of the human cochlea is extended into a 3D geometrical model. Also, spiral computer tomography data of a patient's cochlea is digitally segmented and geometrically represented. The cochlear electrode array is synthesized according to its specification. Then, cochlear implantation is animated with both idealized and real cochlear models. The insertion length, angular position and characteristic frequency of individual electrodes are estimated online during the virtual insertion. The optimization of the processing parameters is done to demonstrate the feasibility of this technology for clinical applications     

Network traffic prediction based on INGARCH model

In this paper, we introduce the integer-valued generalized autoregressive conditional heteroscedasticity (INGARCH) as a network traffic prediction model. As the INGARCH is known as a non-linear analytical model that could capture the characteristics of network traffic such as Poisson packet arrival and long-range dependence property, INGARCH seems to be an adequate model for network traffic prediction. Based on the investigation for the traffic arrival process in various network topologies including IoT and VANET, we could confirm that assuming the Poisson process as packet arrival works for some networks and environments of networks. The prediction model is generated by estimating parameters of the INGARCH process and predicting the Poisson parameters of future-steps ahead process using the conditional maximum likelihood estimation method and prediction procedure, respectively. Its performance is compared with those of three different models; autoregressive integrated moving average, GARCH, and long short-term memory recurrent neural network. Anonymized passive traffic traces provided by the Center for Applied Internet Data Analysis are used in the experiment. Numerical results show that the proposed model predicts better than the three models in terms of measurements used in prediction models. Based on the study, we can conclude the followings: INGARCH can capture the characteristics of network traffic better than other statistic models, it is more tractable than neural networks (NNs) overcoming the black-box nature of NNs, and the performances of some statistical models are comparable or even superior to those of NNs, especially when the data is insufficient to apply deep NNs.

     

Evolutionary fuzzy neural networks for hybrid financial prediction

In this paper, an evolutionary fuzzy neural network using fuzzy logic, neural networks (NNs), and genetic algorithms (GAs) is proposed for financial prediction with hybrid input data sets from different financial domains. A new hybrid iterative evolutionary learning algorithm initializes all parameters and weights in the five-layer fuzzy NN, then uses GA to optimize these parameters, and finally applies the gradient descent learning algorithm to continue the optimization of the parameters. Importantly, GA and the gradient descent learning algorithm are used alternatively in an iterative manner to adjust the parameters until the error is less than the required value. Unlike traditional methods, we not only consider the data of the prediction factor, but also consider the hybrid factors related to the prediction factor. Bank prime loan rate, federal funds rate and discount rate are used as hybrid factors to predict future financial values. The simulation results indicate that hybrid iterative evolutionary learning combining both GA and the gradient descent learning algorithm is more powerful than the previous separate sequential training algorithm described in.     

Improved Performance of Polymer:Polymer Solar Cells by Doping Electron‐Accepting Polymers with an Organosulfonic Acid

The performance of polymer:polymer solar cells that are made using blend films of poly(3‐hexylthiophene) (P3HT) and poly(9,9‐dioctylfluorene‐co‐ benzothiadiazole (F8BT) is improved by doping the F8BT polymer with an organosulfonic acid [4‐ethylbezenesulfonic acid (EBSA)]. The EBSA doping of F8BT, to form F8BT‐EBSA, is performed by means of a two‐stage reaction at room temperature and 60°C with various EBSA weight ratios. The X‐ray photoelectron spectroscopy measurement reveals that both sulfur and nitrogen atoms in the F8BT polymer are affected by the EBSA doping. The F8BT‐EBSA films exhibit huge photoluminescence quenching, ionization potential shift toward lower energy, and greatly enhanced electron mobility. The short‐circuit current density of solar cells is improved by ca. twofold (10 wt.% EBSA doping), while the open‐circuit voltage increases by ca. 0.4 V. Consequently, the power conversion efficiency was improved by ca. threefold, even though the optical density of the P3HT:F8BT‐EBSA blend film is reduced by 10 wt.% EBSA doping due to the nanostructure and surface morphology change.     

Fast adaptive digital equalization by recurrent neural networks

Neural networks (NNs) have been extensively applied to many signal processing problems. In particular, due to their capacity to form complex decision regions, NNs have been successfully used in adaptive equalization of digital communication channels. The mean square error (MSE) criterion, which is usually adopted in neural learning, is not directly related to the minimization of the classification error, i.e., bit error rate (BER), which is of interest in channel equalization. Moreover, common gradient-based learning techniques are often characterized by slow speed of convergence and numerical ill conditioning. In this paper, we introduce a novel approach to learning in recurrent neural networks (RNNs) that exploits the principle of discriminative learning, minimizing an error functional that is a direct measure of the classification error. The proposed method extends to RNNs a technique applied with success to fast learning of feedforward NNs and is based on the descent of the error functional in the space of the linear combinations of the neurons (the neuron space); its main features are higher speed of convergence and better numerical conditioning w.r.t. gradient-based approaches, whereas numerical stability is assured by the use of robust least squares solvers. Experiments regarding the equalization of PAM signals in different transmission channels are described, which demonstrate the effectiveness of the proposed approach     

A quasi-parametric algorithm for synthetic aperture radar target feature extraction and imaging with angle diversity

We present synthetic aperture radar (SAR) target feature extraction and imaging techniques with angle divesity. We first establish a flexible data model that describes each target scatterer as a two-dimensional (2D) complex sequence with arbitrary amplitude and constant phase in range and cross-range. A new algorithm, referred to as the QUasiparametric ALgorithm for target feature Extraction (QUALE), is then presented for SAR target feature extraction via data fusion through angle diversity based on the flexible data model. QUALE first estimates the model parameters, which include, for each scatterer, a 2D arbitrary real-valued amplitude sequence, a constant phase, and scatterer locations in range and cross-reange. QUALE then averages the estimated 2D real-valued amplitude sequence over range by making the assumption that the scatterer radar cross section is approximately consant. QUALE next models the so-obtained 1D sequence with a simple sinc function by assuming that the scatterer is approximately a dihedral (a trihedral is approximated as a very short dihedral) and estimates the relevant sinc function parameters by minimizing a nonlinear least-squares fitting function. Finally, the approximate 2D SAR image is reconstructed by using the estimated features. Numerical examples are given to demonstrate the perfomance of the proposed algorithm.This work was supported in part by AFRL/SNAT, Air Force Research Laboratory, Air Force Materiel Command, USAF, under grant no. F33615-99-1-1507, and the National Science Foundation Grant MIP-9457388. The U.S. Goverment is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon.     

Precipitation retrieval from spaceborne microwave radiometers basedon maximum a a posteriori probability estimation

A retrieval technique for estimating rainfall rate and precipitating cloud parameters from spaceborne multifrequency microwave radiometers is described. The algorithm is based on the maximum a posteriori probability criterion (MAP) applied to a simulated data base of cloud structures and related upward brightness temperatures. The cloud data base is randomly generated by imposing the mean values, the variances, and the correlations among the hydrometeor contents at each layer of the cloud vertical structure, derived from the outputs of a time-dependent microphysical cloud model. The simulated upward brightness temperatures are computed by applying a plane-parallel radiative transfer scheme. Given a multifrequency brightness temperature measurement, the MAP criterion is used to select the most probable cloud structure within the cloud-radiation data base. The algorithm is computationally efficient and has been numerically tested and compared against other methods. Its potential to retrieve rainfall over land has been explored by means of Special Sensor Microwave/Imager measurements for a rainfall event over Central Italy. The comparison of estimated rain rates with available raingauge measurements is also shown     

Object-based analysis-synthesis coding (OBASC) based on the sourcemodel of moving flexible 3-D objects

This correspondence investigates object-based analysis-synthesis coding (OBASC) for the encoding of moving images at very low data rates. According to the source model, each moving object of an image is described and encoded by three parameter sets defining its motion, shape, and surface color. The parameter sets of each object are obtained by model-based image analysis. They are coded by an object-dependent parameter coding. Using the coded parameter sets, an image can be synthesized by model-based image synthesis. Here, OBASC based on the source model of "moving flexible 3-D objects with 3-D motion" (F3D) is introduced. The efficiency of this source model F3D is compared to the efficiency of OBASC based on the source model of "moving rigid 3-D objects with 3-D motion" (R3D). Compared to R3D, F3D requires the additional transmission of flexible-shape parameters. Therefore, the source model F3D is only applied in those areas of the image which cannot be described by the source model R3D. The new source model F3D reduces the bit rate from 64 to 56 kb/s, providing the same picture quality measured by the SNR of the encoded color parameters.     

Multi-factor person entity relation extraction model based on distant supervision

Aiming at the problem that the basic assumption of distant supervision was too strong and easy to produce noise data,a model of the person entity relation extraction which could automatically filter the training data generated by distant supervision was proposed.For training data generation,the data produced by distant supervision would be filtered by multiple instance learning and the method of TF-IDF-based relation keyword detecting,which tried to make the training data has the manual annotation quality.Furthermore,the model combined lexical and syntactic features to extract the effective relation feature vector from two angles of words and semantics for classifier.The experiment results on large scale real-world datasets show that the proposed model outperforms other relation extraction methods which based on distant supervision.     

Retrieval of sea water optically active parameters fromhyperspectral data by means of generalized radial basis function neuralnetworks

The authors present a new methodology for estimating the concentration of sea water optically active constituents from remotely sensed hyperspectral data, based on generalized radial basis function neural networks (GRBF-NNs). This family of NNs is particularly suited to approximate relationships like those between hyperspectral reflectance data and the concentrations of optically active constituents of the water body, which are highly nonlinear, especially in case II waters. Three main water constituents are taken into account: phytoplankton, nonchlorophyllous particles, and yellow substance. Each parameter is estimated by means of a specific multi-input single-output GRBF-NN. The authors adopt a recently proposed network learning strategy based on the combined use of the regression tree procedure and forward selection. The effectiveness of this approach, which is completely general and can be easily applied to any hyperspectral sensor, is proved using data simulated with an ocean color model over the channels of the medium resolution imaging spectrometer (MERIS), the new generation ESA sensor to be launched in 2001. The authors define the estimation algorithms over waters of cases I, II, and I+II and compare their performance with that of classical band-ratio, single-band, and multilinear algorithms. Generally, the GRBF-NN algorithms outperform the classical ones, except for the multilinear over case I waters. A particular improvement Is over case II waters, where the mean square error (MSE) can be reduced by one or two orders of magnitude over the error of multilinear and band-ratio algorithms, respectively     

基于神经网络的某型飞机发动机故障诊断研究

航空发动机故障诊断技术对避免飞行事故和降低飞行器运行成本是十分重要的。提出一种BP网络对某型飞机发动机进行故障诊断,但是由于BP网络收敛速度较慢而且容易陷入局部极小值,特别是BP网络通常只能给出一个解,受训练样本病态影响大。因此通过对BP网络的改进,建立了L-M算法神经网络的飞机发动机故障诊断模型。实验表明,该网络在一定程度上克服了BP网络存在的的问题,在逼近能力、分类能力和学习速度等方面均优于BP网络。为机务人员提供了有效的、科学的发动机故障诊断方法,该种评估手段较好地解决了发动机故障诊断问题,在飞行安全中发挥着越来越大的作用。