基于长短时记忆神经网络的能谱核素识别方法

针对新兴的能谱核素识别方法在混合放射性核素的噪声环境中存在识别速度慢、准确率较低等问题,提出了基于长短时记忆神经网络（LSTM）的能谱核素识别方法。实验使用溴化镧（LaBr₃）晶体探测器,分别对环境中⁶⁰Co、¹³⁷Cs放射性源分组测量得到能谱数据集,首先使用数据平滑方法和归一化方法进行数据预处理,然后将能谱数据按时间序列分组以获得可用的输入序列数组,最后训练LSTM模型得到预测结果。通过基于BP神经网络和卷积神经网络（CNN）的两个能谱识别模型进行对比,得到在测试集中平均识别率分别为83.45%和86.21%,而LSTM能谱识别模型平均识别率为93.04%,实验结果表明,该能谱模型在核素识别效果中表现较好,可用于快速的能谱核素识别设备上。     

Ionospheric vertical total electron content prediction model in low-latitude regions based on long short-term memory neural network

Ionosphere delay is one of the main sources of noise affecting global navigation satellite systems, operation of radio detection and ranging systems and very-long-baseline-interferometry. One of the most important and common methods to reduce this phase delay is to establish accurate nowcasting and forecasting ionospheric total electron content models. For forecasting models, compared to mid-to-high latitudes, at low latitudes, an active ionosphere leads to extreme differences between long-term prediction models and the actual state of the ionosphere. To solve the problem of low accuracy for long-term prediction models at low latitudes, this article provides a low-latitude, long-term ionospheric prediction model based on a multi-input-multi-output, long-short-term memory neural network. To verify the feasibility of the model, we first made predictions of the vertical total electron content data 24 and 48 hours in advance for each day of July 2020 and then compared both the predictions corresponding to a given day, for all days. Furthermore, in the model modification part, we selected historical data from June 2020 for the validation set, determined a large offset from the results that were predicted to be active, and used the ratio of the mean absolute error of the detected results to that of the predicted results as a correction coefficient to modify our multi-input-multi-output long short-term memory model. The average root mean square error of the 24-hour-advance predictions of our modified model was 4.4 TECU, which was lower and better than 5.1 TECU of the multi-input-multi-output, long short-term memory model and 5.9 TECU of the IRI-2016 model.     

卷积神经网络主动目标方位估计

复杂海洋环境中信道的传输特性、时空变化、频散效应等一定程度上制约了主动声呐目标方位估计的性能。该文引入卷积神经网络(CNN),提出了适用于主动声呐中目标方位的高精度估计方法。仿真声场环境为浅海负梯度,主动发射信号为具有多普勒不变性质的双曲调频信号,水平线列阵作为接收装置,目标按仿真路线运动。该文利用Kraken进行声场数据仿真,并对接收的信号在频域做均匀加权常规波束形成,进而进行卷积神经网络的模型训练和测试。数值仿真研究表明,该文所用方法可以有效估计目标波达方向,对信噪比具有一定的鲁棒性。     

Preliminary abnormal electrocardiogram segment screening method for Holter data based on long short-term memory networks

Holter usually monitors electrocardiogram(ECG)signals for more than 24 hours to capture short-lived cardiac abnormalities.In view of the large amount of Holter data and the fact that the normal part accounts for the majority,it is reasonable to design an algorithm that can automatically eliminate normal data segments as much as possible without missing any abnormal data segments,and then take the left segments to the doctors or the computer programs for further diagnosis.In this paper,we propose a preliminary abnormal segment screening method for Holter data.Based on long short-term memory(LSTM)networks,the prediction model is established and trained with the normal data of a monitored object.Then,on the basis of kernel density estimation,we learn the distribution law of prediction errors after applying the trained LSTM model to the regular data.Based on these,the preliminary abnormal ECG segment screening analysis is carried out without R wave detection.Experiments on the MIT-BIH arrhythmia database show that,under the condition of ensuring that no abnormal point is missed,53.89% of normal segments can be effectively obviated.This work can greatly reduce the workload of subsequent further processing.     

基于长短期记忆网络与特征融合的微表情识别算法

为了提高视频识别领域中微表情识别的准确率,提出了一种基于长短期记忆网络与特征融合的微表情识别算法。提取微表情图像的颜色特征和纹理特征,将所提取的空间特征传入卷积神经网络进行融合。设计了学习时域相关性的长短期记忆网络结构,将融合的特征集传入长短期记忆网络学习微表情的时域特征,将长短期记忆网络接入分类器网络识别出微表情的类标签。在两个公开的微表情识别数据集上完成了验证实验,结果显示算法实现了较好的微表情识别性能,在SMIC数据集和CASMEⅡ数据集上的准确率分别达到64.7%和65.8%.     

局域互联神经网络的关联存储

基于全局互联的Hopfield模型,本文提出了局域互联关联存储的新概念.与全局互联相比,局域互联具有较小的关联矩阵,因而,有利于用现有的空间光调制器加以实现.同时,计算机模拟结果表明,它仍然具有全局关联存储的能力.     

Associative memory on a small-world neural network

We study a model of associative memory based on a neural network with small-world structure. The efficacy of the network to retrieve one of the stored patterns exhibits a phase transition at a finite value of the disorder. The more ordered networks are unable to recover the patterns, and are always attracted to non-symmetric mixture states. Besides, for a range of the number of stored patterns, the efficacy has a maximum at an intermediate value of the disorder. We also give a statistical characterization of the spurious attractors for all values of the disorder of the network.Received: 12 January 2004, Published online: 28 May 2004PACS: 84.35. + i Neural networks - 89.75.Hc Networks and genealogical trees - 87.18.Sn Neural networks     

两维局域互联神经网络的关联存储

本文将局域互联神经网络的新概念推广到两维情形,并对两维局域互联关联存储进行了理论分析和大量的计算机模拟.结果表明,两维局域互联神经网络的优点是,在满足存储容量限制的前提下,它与全局互联神经网络具有相同的关联存储能力,而其互联权重矩阵要比全局互联网络小得多.因而,有利于使用现有的空间光调制器实现两维大规模的人工神经网络.     

基于注意力机制与图卷积神经网络的单目红外图像深度估计

对场景中的物体进行深度估计是无人驾驶领域中的关键问题,红外图像有利于在光线不佳的情况下解决深度估计问题.针对红外图像纹理不清晰与边缘信息不丰富的特点,提出了将注意力机制与图卷积神经网络相结合来解决单目红外图像深度估计问题.首先,在深度估计问题中,图像中每个像素点的深度信息不仅与其周围像素点的深度信息相关,还需考虑更大范...     

基于神经网络的视觉系统标定方法

为了解决摄像机标定存在的若干问题 ,根据立体视觉原理 ,提出了基于神经网络的双目视觉系统标定方法。通过对双目摄像机的有效视场分析 ,确定了一次测量面积 ,并把像对视差作为网络输入 ,建立空间点世界坐标与图像坐标非线性映射关系 ,使系统不经过复杂的摄像机内外参数标定 ,就能直接提取物体的三维信息 ,增加了系统的灵活性。实验证明 ,该方法有效可行     

基于神经网络的低照度彩色图像增强算法

由于低照度彩色图像存在整体亮度低、对比度低、颜色偏暗和信噪比低等特点,所以经典图像增强算法对其增强效果非常有限。提出了一种利用BP神经网络进行彩色图像增强的算法,并将RGB图像转换成HSI图像,以保证增强处理不引起图像的色彩失真。实验证明:该方法显著地改善了低照度彩色图像的视觉效果,提高了图像整体亮度和图像的信噪比,可调节图像的动态范围,能增强图像的对比度和细节,可增加图像信息熵。     

Solder joint inspection based on neural network combined with genetic algorithm

To improve the performance of automatic optical inspection (AOI), a neural network combined with genetic algorithm for the diagnosis of solder joint defects on printed circuit boards (PCBs) assembled in surface mounting technology (SMT) is presented. Six types of solder joint have been classified in respect to the reality in the manufacture. The images of solder joint under test are acquired and 14 features are extracted as input features for the classification. The neural network is easily become over-fitting because these input features are not independent of each other, so the genetic algorithm is introduced to select and remove redundant input features. The experimental results have proved that the neural network combined with genetic algorithm reduced the number of input feature and had a satisfying recognition rate.     

Model architecture for associative memory in a neural network of spiking neurons

A synaptic connectivity model is assembled on a spiking neuron network aiming to build up a dynamic pattern recognition system. The connection architecture includes gap junctions and both inhibitory and excitatory chemical synapses based on Hebb’s hypothesis. The network evolution resulting from external stimulus is sampled in a properly defined frequency space. Neurons’ responses to different current injections are mapped onto a subspace using Principal Component Analysis. Departing from the base attractor, related to a quiescent state, different external stimuli drive the network to different fixed points through specific trajectories in this subspace.     

State of charge estimation of Li-ion batteries in an electric vehicle based on a radial-basis-function neural network

The on-line estimation of the state of charge(SOC) of the batteries is important for the reliable running of the pure electric vehicle in practice.Because a nonlinear feature exists in the batteries and the radial-basis-function neural network(RBF NN) has good characteristics to solve the nonlinear problem,a practical method for the SOC estimation of batteries based on the RBF NN with a small number of input variables and a simplified structure is proposed.Firstly,in this paper,the model of on-line SOC estimation with the RBF NN is set.Secondly,four important factors for estimating the SOC are confirmed based on the contribution analysis method,which simplifies the input variables of the RBF NN and enhances the real-time performance of estimation.Finally,the pure electric buses with LiFePO 4 Li-ion batteries running during the period of the 2010 Shanghai World Expo are considered as the experimental object.The performance of the SOC estimation is validated and evaluated by the battery data from the electric vehicle.     

神经网络技术在激光脉冲解码领域的应用研究

为提高激光脉冲解码过程的准确性和识别效率,采用神经网络技术对激光脉冲编码解码进行了仿真研究.应用线性神经网络对有规律的编码,如周期型编码和等差型编码,进行了识别.仿真结果表明,对于PCM码,需要约2个周期的脉冲就可准确预测下一个脉冲到达的时间;对于等差型码,需要4个脉冲就可以达到精度要求.然后,应用概率神经网络对伪随机型编码的最小周期进行了识别.仿真结果表明,可以在信息量较少的情况下准确识别此类型编码的最小周期.     

色貌模型的人工神经网络方法的研究

色貌模型(CAM)主要解决不同观察条件、不同背景和不同环境下的颜色真实再现问题。采用人工神经网络(ANN)的方法来实现目前最新的色貌模型CIECAM02的预测,包括正向预测(从色度参数到色貌属性参数)和逆向预测(从色貌属性参数到色度参数),应用自然色系统(NCS)中的部分色样作为神经网络的训练和测试样本。由于正向输出色貌属性参数空间不是均匀的,对于网络预测精度用特殊方法评估,而对于逆向模型则可直接利用LAB色差公式评价。测试的结果表明:用神经网络对CIECAM02模型的预测达到了较高的精度。     

基于神经网络模型的CCD像差修正

针对传统CCD标定中由于修正像差造成标定算法复杂、精度不高、结果不稳定的特点,提出了用4层前向神经网络模型修正CCD像差的方法。为了检验算法,用修正像差后的CCD对空间点进行三维重建,并与传统方法作了对比。结果表明,基于神经网络模型的CCD像差修正方法可以获得比传统方法更高的精度和稳定性。     

基于混沌神经网络的单向Hash函数

提出了一种基于混沌神经网络的单向Hash函数，该方法通过使用以混沌分段线性函数作为输出函数的神经网络和基于时空混沌的密钥生成函数实现明文和密钥信息的混淆和扩散，并基于密码块连接模式实现对任意长度的明文序列产生128位的Hash值.理论分析和实验结果表明，提出的Hash函数可满足所要求的单向性，初值和密钥敏感性，抗碰撞性和实时性等要求. 关键词： 混沌神经网络 Hash函数 分段线性混沌映射 时空混沌     

基于BP神经网络的焊点自动光学检测

提出了一种新的表面组装焊点的自动光学检测分类方法。采用单层环形光源获取焊点图像,并根据归一化分割曲线方程将焊点图像分成四部分,分别提取其特征。使用BP神经网络将焊点按照锡量多少分成三类:少锡,容许和多锡。实验结果证明该方法分类正确率达到了99.2%,具有较高的实用价值。     

基于BP神经网络的闪光照相图像复原

针对闪光照相系统模糊较大、成像信噪比较低的问题,提出了一种基于BP神经网络的闪光照相图像复原方法。该方法利用BP神经网络的泛化能力,用样本图像对网络进行训练,建立退化图像与真实图像之间的非线性映射关系,然后将待复原图像分区,利用训练好的BP神经网络对待复原图像的边界区域进行复原处理。数值试验表明,在系统点扩展函数未知的情况下,该算法能较好再现图像边缘信息,复原出的图像在信噪比和视觉方面都有较大提高。