Beam-filling effect correction with subpixel cloud fraction using a neural network

In this study, the effects of cloud inhomogeneity on microwave rain rate retrievals are investigated. A single-channel (85 GHz) empirically based algorithm using a neural network approach is presented. The objective is to correct the beam-filling error (BFE), that might occur because of the inherent variability within coarse microwave pixels, with subpixel information. To this aim, we used the Tropical Rainfall Measuring Mission passive microwave, thermal infrared and radar data. The integration of spatial information into the retrieval algorithm enables us to partially overcome the BFE. We use two parameters which characterize the horizontal cloud inhomogeneity within the microwave radiometer field of view, and we add them to simulated brightness temperatures as inputs of the neural network algorithm. The first one is the cloud fraction derived from infrared measurement, and the second corresponds to the fraction of the rainy area derived from radar measurements. The output rain rates were validated using the Precipitation Radar data. It was found that adding cloud fraction of microwave pixels, can lead to more accurate retrievals. Instantaneous precipitation estimates demonstrated correlations of /spl sim/0.6-0.7 and /spl sim/0.7-0.8 with radar-derived rain rates, for ocean and land retrievals respectively. In spite of the problem inherent in deriving the cloud (or rain) fraction, the initial validation results presented in this study are reasonably encouraging and show the advantage of utilizing the information from different sensors in order to optimize the retrieval of rainfall.     

TBNN: totally-binary neural network for image classification

Most binary networks apply full precision convolution at the first layer. Changing the first layer to the binary convolution will result in a significant loss of accuracy. In this paper, we propose a new approach to solve this problem by widening the data channel to reduce the information loss of the first convolutional input through the sign function. In addition, widening the channel increases the computation of the first convolution layer, and the problem is solved by using group convolution. The experimental results show that the accuracy of applying this paper''s method to state-of-the-art (SOTA) binarization method is significantly improved, proving that this paper''s method is effective and feasible.     

A neural network approach to MVDR beamforming problem

A Hopfield-type neural network approach which leads to an analog circuit for implementing the real-time adaptive antenna array is presented. An optimal array pattern can be steered by updating the weights across the array in order to maximize the output signal-to-noise ratio (SNR). The problem of adjusting the array weights can be characterized as a constrained quadratic nonlinear programming. The adjustment of settings is required to respond to a rapid time-varying environment. A Hopfield-type neural net with a number of graded-response neurons designed to perform the constrained quadratic nonlinear programming would lead to a solution in a time determined by RC time constants, not by algorithmic time complexity. The constrained quadratic programming neural net has associated it with an energy function which the net always seeks to minimize. A fourth-order Runge-Kutta simulation shows that the circuit operates at a much higher speed than conventional techniques and the computation time of solving a linear array of 10 elements is about 0.1 ns for RC=5×10 ^-9     

A NEURAL NETWORK APPROACH TO GATE MATRIX LAYOUT

Gate matrix layout problem plays an important role in integrated circuit design, but its optimization is NP-hard. In this paper, typical gate layout problem is analysed and adapted to neural network representation, furthermore the simulated results are given.     

A pyramid approach to subpixel registration based on intensity

We present an automatic subpixel registration algorithm that minimizes the mean square intensity difference between a reference and a test data set, which can be either images (two-dimensional) or volumes (three-dimensional). It uses an explicit spline representation of the images in conjunction with spline processing, and is based on a coarse-to-fine iterative strategy (pyramid approach). The minimization is performed according to a new variation (ML*) of the Marquardt-Levenberg algorithm for nonlinear least-square optimization. The geometric deformation model is a global three-dimensional (3-D) affine transformation that can be optionally restricted to rigid-body motion (rotation and translation), combined with isometric scaling. It also includes an optional adjustment of image contrast differences. We obtain excellent results for the registration of intramodality positron emission tomography (PET) and functional magnetic resonance imaging (fMRI) data. We conclude that the multiresolution refinement strategy is more robust than a comparable single-stage method, being less likely to be trapped into a false local optimum. In addition, our improved version of the Marquardt-Levenberg algorithm is faster.     

Natural frequency-based neural network approach to radar target recognition

Time domain response-based neural networks and frequency domain response-based neural networks have been proposed for radar target recognition. We propose a natural frequency-based neural network for radar target recognition. Our scheme takes advantage of an aspect angle independence of a natural frequency. It is shown from experimental results that a natural frequency based-neural network using the first natural frequency pair is superior to a time domain response-based neural network in the case of a single aspect angle and that a natural frequency based-neural network using the first natural frequency pair or the first two natural frequency pairs is superior to a time domain response-based neural network in the case of a multiple aspect angle.     

A knowledge-based approach to improve neural network control design

Control engineers have been investigating and developing different on-line adaptation schemes to fine-tune performance of controllers after off-line design. Artificial neural network technology has shown satisfactory results for many control applications. However, certain types of nonlinear problems are difficult for the neural controller to learn by conventional on-line adaptation schemes, while the nonlinear system can be effectively controlled by incorporating heuristics knowledge. This paper presents an effective approach to incorporate heuristics control knowledge into a neural controller by off-line pre-training, then fine-tune the neural controller performance further by on-line adaptation. Experimental results on a servomotor system with significant nonlinear friction characteristics are used to demonstrate the effectiveness of the design approach.     

Residual life predictions from vibration-based degradation signals: a neural network approach

Maintenance of mechanical and rotational equipment often includes bearing inspection and/or replacement. Thus, it is important to identify current as well as future conditions of bearings to avoid unexpected failure. Most published research in this area is focused on diagnosing bearing faults. In contrast, this paper develops neural-network-based models for predicting bearing failures. An experimental setup is developed to perform accelerated bearing tests where vibration information is collected from a number of bearings that are run until failure. This information is then used to train neural network models on predicting bearing operating times. Vibration data from a set of validation bearings are then applied to these network models. Resulting predictions are then used to estimate the bearing failure time. These predictions are then compared with the actual lives of the validation bearings and errors are computed to evaluate the effectiveness of each model. For the best model, we find that 64% of predictions are within 10% of actual bearing life, while 92% of predictions are within 20% of the actual life.     

Solder joints inspection using a neural network and fuzzyrule-based classification method

In this paper, we described an approach in automation, the visual inspection of solder joint defects of surface mounted components on a printed circuit board, using a neural network with fuzzy rule-based classification method. Inherently, the solder joints have a curved, tiny, and specular reflective surface. This presents the difficulty in taking good images of the solder joints. Furthermore, the shapes of the solder joints tend to greatly vary with their soldering conditions, and are not identical with each other, even though some of the solder joints belong to a set of the same soldering quality. This problem makes it difficult to classify the solder joints according to their properties. To solve this intricate problem, a new classification method is here proposed which consists of two modules: one based upon an unsupervised neural network, and the other based upon a fuzzy set theory. The novel idea of this approach is that a fuzzy rule table reflecting the knowledge of criteria of a human inspector, is utilized in order to correct any possible misclassification made by the neural network module. The performance of the proposed approach was tested on numerous samples of printed circuit boards in commercially available computers, and then compared with that of a human inspector. Experimental results reveal that the proposed method is superior to the neural network classification method alone, in terms of its accuracy of classification     

A neural network approach to MMSE receivers in a DS-CDMA multipath fading environment

This letter deals with an advanced minimum mean-squared error receiver for applications to uplink transmissions in a multiuser code-division multiple-access system. The receiver is implemented by means of a suitable neural network in order to enhance the receiver convergence speed in the case of fast fading. Performance comparisons with classical approaches highlights a better behavior for the proposed scheme.     

Comments on a “A neural network approach to MVDR beamformingproblem”

Several corrections to the above-titled paper by P.R. Chang et al. (ibid, vol. 40, pp 313-322, 1992) are given by the commenter, who also discusses a number of Chang et al.'s results. The commenter also discusses the practical issues that arise when an array operates in a nonstationary environment     

一种基于卷积神经网络的雷达目标分类方法

雷达作为对低空和地面目标探测及监视预警的主要手段,在安全领域应用广泛。针对现阶段实际应用中雷达目标分类技术中过于依赖人工提取特征的问题,提出了一种基于卷积神经网络的分类方法,对雷达回波数据进行二维傅里叶变换得到距离-多普勒图像,再以距离-多普勒图集作为数据集,训练神经网络,得到能够完成雷达目标识别的网络模型。结果表明,相较于传统方法,基于卷积神经网络的目标识别模型在省去人工工作的同时提高了目标识别精度。     

基于BP神经网络的音乐分类模型

针对单一特征难以建立理想音乐分类模型的不足,为了帮助用户找到自己喜欢的音乐,提出BP神经网络的音乐分类模型。首先提取音乐的多种类型特征,便于对音乐信息进行准确描述,然后将这些特征组合在一起作为音乐分类模型的输入向量,通过BP神经网络的智能学习建立音乐分类模型,最后在Matlab 2016平台下进行多个音乐分类实验。结果表明,该模型克服了单一特征提供信息简单的局限性,提高了音乐的分类正确率,而且音乐分类的实时性较好,可以用于网络上的音乐检索研究。     

A neural network modeling approach to circuit optimization andstatistical design

The trend of using accurate models such as physics-based FET models, coupled with the demand for yield optimization results in a computationally challenging task. This paper presents a new approach to microwave circuit optimization and statistical design featuring neural network models at either device or circuit levels. At the device level, the neural network represents a physics-oriented FET model yet without the need to solve device physics equations repeatedly during optimization. At the circuit level, the neural network speeds up optimization by replacing repeated circuit simulations. This method is faster than direct optimization of original device and circuit models. Compared to existing polynomial or table look-up models used in analysis and optimization, the proposed approach has the capability to handle high-dimensional and highly nonlinear problems     

A neural network approach to real-time condition monitoring ofinduction motors

A neural network-based incipient fault detector for small and medium-size induction motors is developed. The detector avoids the problems associated with traditional incipient fault detection schemes by employing more readily available information such as rotor speed and stator current. The neural network design is evaluated in real time in the laboratory on a 3/4 hp permanent magnet induction motor. The results of this evaluation indicate that the neural-network-based incipient fault detector provides a satisfactory level of accuracy, greater than 95%, which is suitable for real-world applications     

A modular neural network approach to microelectronic circuit yield optimization

The unpredictable variation in microelectronic circuits due to process tolerances increases significantly with increased levels of miniaturization. If ignored, the variation will result in poor manufacturing yield. If a worst-case approach is adopted, a loss of competitive edge results. This situation provides the motivation for efficient robust design of VLSI circuits, the subject of this paper. Given the need for efficiency of analysis without significant loss of accuracy, a method is proposed which generates a neural network for mapping process-level parameters to circuit performance. The approach uses a modular neural network—an adaptive mixture of local experts competing to learn different aspects of a problem. Once the neural network model is established and validated, it is employed in performing extremely efficient optimization of the circuit yield at minimal cost: the trained ANN acts as a cheap but accurate simulator which when supplied with a set of inputs which characterize transistors at the process and device level calculates the circuit performance with 97% accuracy at 1% of the cost of a full SPICE simulation. Even when the cost of ANN training is factored in, average cost savings of 80% are achieved during yield optimization. The neural net approach offers significant advantages including vastly reduced computational cost with little loss of accuracy and complete generality of application.     

A neural network approach to estimating rainfall from spacebornemicrowave data

Various techniques use microwave (MW) brightness temperature (BT) data, obtained from remote sensing orbiting platforms, to calculate rain rates. The most commonly used techniques are based on regressions or other statistical methods. An emerging tool in rainfall estimation using satellite data is artificial neural networks (NNs), NNs are mathematical models that are capable of learning complex relationships. They consist of highly interconnected, interactive data processing units. NNs are implemented in this study to estimate rainfall, and backpropagation is used as a learning scheme. The inputs for the training phase are BTs and the outputs are rainfall rates, all generated by three-dimensional (3D) simulations based on a 3D stochastic, space-time rainfall model, and a 3D radiative transfer model. Once training is complete the NNs are presented with multi-frequency and polarized (horizontal and vertical) BT data, obtained from the Special Sensor Microwave/Imager (SSM/I) instrument onboard the F10 and F11 polar-orbiting meteorological satellites. Hence, rainrates corresponding to real BT measurements are generated. The rainfall rates are also estimated using a log-linear regression model. Comparison of the two approaches, using simulated data, shows that the NN can represent more accurately the underlying relationship between BT and rainrate than the regression model, Comparison of the rates, estimated by both methods, with radar-estimated rainrates shows that NNs outperform the regression model. This study demonstrates the great potential of NNs in estimating rainfall from remotely sensed data     

基于卷积神经网络和胶囊神经网络的宫颈病变图像分类方法研究

为了更好地治疗宫颈癌,准确确定患者的宫颈类型是至关重要的。因此,用于检测和划分宫颈类型的自动化方法在该领域中具有重要的医学应用。虽然深度卷积神经网络和传统的机器学习方法在宫颈病变图像分类方面已经取得了良好的效果,但它们无法充分利用图像和图像标签的某些关键特征之间的长期依赖关系。为了解决这个问题,文章引入了胶囊网络(CapsNet),将CNN和CapsNet结合起来,以提出CNN-CapsNet框架,该框架可以加深对图像内容的理解,学习图像的结构化特征,并开展医学图像分析中大数据的端到端训练。特别是,文章应用迁移学习方法将在ImageNet数据集上预先训练的权重参数传输到CNN部分,并采用自定义损失函数,以便网络能够更快地训练和收敛,并具有更准确的权重参数。实验结果表明,与ResNet和InceptionV3等其他CNN模型相比,文章提出的网络模型在宫颈病变图像分类方面更加准确、有效。     

A neural network algorithm for sea ice edge classification

The NASA Scatterometer (NSCAT), launched in August 1995, is designed to measure wind vectors over ice-free oceans. To prevent contamination of the wind measurements, by the presence of sea ice, algorithms based on neural network technology have been developed to classify ice-free ocean surfaces. Neural networks trained using polarized alone and polarized plus multi-azimuth “look” Ku-band backscatter are described. Algorithm skill in locating the sea ice edge around Antarctica is experimentally evaluated using backscatter data from the Seasat-A Satellite Scatterometer that operated in 1978. Comparisons between the algorithms demonstrate a slight advantage of combined polarization and multi-look over using co-polarized backscatter alone. Classification skill is evaluated by comparisons with surface truth (sea ice maps), subjective ice classification, and independent over lapping scatterometer measurements (consecutive revolutions)