A one-class peeling method for multivariate outlier detection with applications in phase I SPC

In phase I of statistical process control (SPC), control charts are often used as outlier detection methods to assess process stability. Many of these methods require estimation of the covariance matrix, are computationally infeasible, or have not been studied when the dimension of the data, $p$, is large. We propose the one-class peeling (OCP) method, a flexible framework that combines statistical and machine learning methods to detect multiple outliers in multivariate data. The OCP method can be applied to phase I of SPC, does not require covariance estimation, and is well suited to high-dimensional data sets with a high percentage of outliers. Our empirical evaluation suggests that the OCP method performs well in high dimensions and is computationally more efficient and robust than existing methodologies. We motivate and illustrate the use of the OCP method in a phase I SPC application on a $N=354$, $p=1917$ dimensional data set containing Wikipedia search results for National Football League (NFL) players, teams, coaches, and managers. The example data set and R functions, OCP.R and OCPLimit.R, to compute the respective OCP distances and thresholds are available in the supplementary materials.     

基于卷积神经网络的葡萄叶片氮含量识别方法

为实现对葡萄叶片氮素含量快速、便捷的识别，在卷积神经网络VGG-16网络结构基础上，将数据增广后的图像按不同梯度划分进行模型训练，通过十折交叉验证法探究最佳的训练集与验证集分配比例，并构建4个不同深度的网络模型进行训练对比，采用全局平均池化代替全连接层约简网络参数量。训练结果表明，氮含量梯度设为0.70%、0.35%和0.175%时，室内简单背景识别准确率分别为85.9%、76.2%和71.1%；晴天室外复杂背景下识别准确率分别为44.6%、35.0%和30.4%。研究结果表明利用VGG-16建立的网络学习模型对葡萄叶片氮含量识别提供了一种新的便捷方法，对农业信息化和智能化技术应用具有一定促进作用。     

A three-stage graphics processing unit-based finite element analyses matrix generation strategy for unstructured meshes

With the development of parallel computing architectures, larger and more complex finite element analyses (FEA) are being performed with higher accuracy and smaller execution times. Graphics processing units (GPUs) are one of the major contributors of this computational breakthrough. This work presents a three-stage GPU-based FEA matrix generation strategy with the key idea of decoupling the computation of global matrix indices and values by use of a novel data structure referred to as the neighbor matrix. The first stage computes the neighbor matrix on the GPU based on the unstructured mesh. Using this neighbor matrix, the indices and values of the global matrix are computed separately in the second and third stages. The neighbor matrix is computed for three different element types. Two versions for performing numerical integration and assembly in the same or separate kernels are implemented and simulations are run for different mesh sizes having up to three million degrees of freedom on a single GPU. Comparison with GPU-based parallel implementation from the literature reveals speedup ranging from 4× to 6× for the proposed workload division strategy. Furthermore, the same kernel implementation is found to outperform the separate kernel implementation by 70% to 150% for different element types.     

Electroencephalogram (EEG) Brain Signals to Detect Alcoholism Based on Deep Learning

The detection of alcoholism is of great importance due to its effects on individuals and society. Automatic alcoholism detection system (AADS) based on electroencephalogram (EEG) signals is effective, but the design of a robust AADS is a challenging problem. AADS’ current designs are based on conventional, hand-engineered methods and restricted performance. Driven by the excellent deep learning (DL) success in many recognition tasks, we implement an AAD system based on EEG signals using DL. A DL model requires huge number of learnable parameters and also needs a large dataset of EEG signals for training which is not easy to obtain for the AAD problem. In order to solve this problem, we propose a multi-channel Pyramidal neural convolutional (MP-CNN) network that requires a less number of learnable parameters. Using the deep CNN model, we build an AAD system to detect from EEG signal segments whether the subject is alcoholic or normal. We validate the robustness and effectiveness of proposed AADS using KDD, a benchmark dataset for alcoholism detection problem. In order to find the brain region that contributes significant role in AAD, we investigated the effects of selected 19 EEG channels (SC-19), those from the whole brain (ALL-61), and 05 brain regions, i.e., TEMP, OCCIP, CENT, FRONT, and PERI. The results show that SC-19 contributes significant role in AAD with the accuracy of 100%. The comparison reveals that the state-of-the-art systems are outperformed by the AADS. The proposed AADS will be useful in medical diagnosis research and health care systems.     

Automatic monitoring system for individual dairy cows based on a deep learning framework that provides identification via body parts and estimation of body condition score

Body condition score (BCS) is a common tool for indirectly estimating the mobilization of energy reserves in the fat and muscle of cattle that meets the requirements of animal welfare and precision livestock farming for the effective monitoring of individual animals. However, previous studies on automatic BCS systems have used manual scoring for data collection, and traditional image extraction methods have limited model performance accuracy. In addition, the radio frequency identification device system commonly used in ranching has the disadvantages of misreadings and damage to bovine bodies. Therefore, the aim of this research was to develop and validate an automatic system for identifying individuals and assessing BCS using a deep learning framework. This work developed a linear regression model of BCS using ultrasound backfat thickness to determine BCS for training sets and tested a system based on convolutional neural networks with 3 channels, including depth, gray, and phase congruency, to analyze the back images of 686 cows. After we performed an analysis of image model performance, online verification was used to evaluate the accuracy and precision of the system. The results showed that the selected linear regression model had a high coefficient of determination value (0.976), and the correlation coefficient between manual BCS and ultrasonic BCS was 0.94. Although the overall accuracy of the BCS estimations was high (0.45, 0.77, and 0.98 within 0, 0.25, and 0.5 unit, respectively), the validation for actual BCS ranging from 3.25 to 3.5 was weak (the F1 scores were only 0.6 and 0.57, respectively, within the 0.25-unit range). Overall, individual identification and BCS assessment performed well in the online measurement, with accuracies of 0.937 and 0.409, respectively. A system for individual identification and BCS assessment was developed, and a convolutional neural network using depth, gray, and phase congruency channels to interpret image features exhibited advantages for monitoring thin cows.     

改进的卷积神经网络单幅图像超分辨率重建

针对经典的基于卷积神经网络的单幅图像超分辨率重建方法网络较浅、提取的特征少、重建图像模糊等问题，提出了一种改进的卷积神经网络的单幅图像超分辨率重建方法，设计了由密集残差网络和反卷积网络组成的新型深度卷积神经网络结构。原始低分辨率图像输入网络，利用密集残差学习网络获取更丰富的有效特征并加快特征梯度流动，其次通过反卷积层将图像特征上采样到目标图像大小，再利用密集残差学习高维特征，最后融合不同卷积核提取的特征得到最终的重建图像。在Set5和Set14数据集上进行了实验，并和Bicubic、K-SVD、SelfEx、SRCNN等经典重建方法进行了对比，重建出的图像在整体清晰度和边缘锐度方面更好，另外峰值信噪比（PSNR）平均分别提高了2.69?dB、1.68?dB、0.74?dB和0.61?dB。实验结果表明，该方法能够获取更丰富的细节信息，得到更好的视觉效果，达到了图像超分辨率的增强任务。     

一种基于雷达图表示的数值型数据的CNN分类方法

卷积神经网络（convolutional neural networks， CNN）是一种广泛用于分析视觉图像的分类方法.由于数值数据存在着非线性、耦合性等复杂的空间关系，因此基于CNN的数值型数据的研究较少.本文的目的是找到一种可行的方法，将CNN的应用领域扩展到数值数据.于是提出了一种基于雷达图表示的数值型数据的CNN分类方法（Radar-CNN）.该算法首先将数值数据表示成雷达图形式，然后将其输入CNN中构建分类模型.为了进一步研究特征尺度和序列对性能的影响，提出了两种改进算法Rank Radar-CNN和SFS Radar-CNN.为了验证所提算法的有效性，引入TE化工过程数据集进行实验测试并比较，实验结果表明Radar-CNN及其改进算法具有优异的性能.     

Space Efficient Quantization for Deep Convolutional Neural Networks

Deep convolutional neural networks (DCNNs) have shown outstanding performance in the fields of computer vision, natural language processing, and complex system analysis. With the improvement of performance with deeper layers, DCNNs incur higher computational complexity and larger storage requirement, making it extremely difficult to deploy DCNNs on resource-limited embedded systems (such as mobile devices or Internet of Things devices). Network quantization efficiently reduces storage space required by DCNNs. However, the performance of DCNNs often drops rapidly as the quantization bit reduces. In this article, we propose a space efficient quantization scheme which uses eight or less bits to represent the original 32-bit weights. We adopt singular value decomposition (SVD) method to decrease the parameter size of fully-connected layers for further compression. Additionally, we propose a weight clipping method based on dynamic boundary to improve the performance when using lower precision. Experimental results demonstrate that our approach can achieve up to approximately 14x compression while preserving almost the same accuracy compared with the full-precision models. The proposed weight clipping method can also significantly improve the performance of DCNNs when lower precision is required.

     

自然光照条件下植被几何光学四分量的提取算法

目的 几何光学四分量是指在太阳光照条件下传感器所能观测的4个光学分量，即光照植被、光照土壤、阴影植被和阴影土壤。四分量是构成遥感几何光学模型的重要内容。在近地表遥感应用中，相机俯视拍照是提取四分量的一个途径。准确快速地从图像数据中提取四分量对植被冠层结构参数反演和植被长势监测具有重要意义。方法 植被与土壤二分量的识别是四分量提取的基础。目前大多数二分类算法在自然光照条件复杂时分类误差较大。本文基于卷积神经网络（CNN）和阈值法实现了多种二分类和四分量提取算法。阈值法中，使用SHAR-LABFVC （shadow-resistant algorithm：LABFVC）实现植被与土壤的二分类，并在此基础上应用二次阈值分割获取四分量，称为二次阈值法；基于CNN的方法中，采用U-Net架构，并使用RGB和RGBV数据进行训练得到U-Net和U-Net-V模型，前者完成二分类和四分量任务，后者只完成四分量提取实验。最后，对一种结合U-Net与阈值法的混合算法进行四分量提取实验。结果 本文在18幅图像（1 800个子图）数据上进行了实验，结果表明，与目视解译得到的四分量真值相比较，U-Net-V和混合法精度最高，具有相近的均方根误差（RMSE）（0.06和0.07）和相关系数（0.95和0.94）；二次阈值法与U-Net模型精度略低于上述两种算法，RMSE分别是0.08和0.09，相关系数均为0.88。在二分类实验中，U-Net的分类正确率是91%，SHAR-LABFVC为85%。结论 通过对比实验表明，在二分类问题中，U-Net可以更好地应对复杂自然光照条件下的数字图像。在四分量提取实验中，混合法和U-Net-V的结果优于U-Net与二次阈值法，可以用于提取四分量。     

Object detection based on the multiscale location Enhancement network

For the target detection task,there are two problems in the one-stage network structure of the deep neural network model.First,whether the design of the anchor box hyperparameter is suitable or not will affect the training results of the whole network;second,a large down sampling factor will affect the positioning ability of the target.To solve these problems,this paper proposes a multi-location enhancement network.The structure of the one-stage network model is redesigned,and a better scheme for selecting the super parameters of the anchor frame is proposed.So the efficiency of the first stage network is ensured and the positioning accuracy is better than the previous one.A large number of experiments show that the multi-location enhancement network can achieve a higher positioning accuracy while ensuring real-time performance.The average accuracy of 82.5 is achieved on the public dataset (Pascal VOC 2007).