Sika Deer Behavior Recognition Based on Machine Vision

With the increasing intensive and large-scale development of the sika deer breeding industry, it is crucial to assess the health status of the sika deer by monitoring their behaviours. A machine vision–based method for the behaviour recognition of sika deer is proposed in this paper. Google Inception Net (GoogLeNet) is used to optimise the model in this paper. First, the number of layers and size of the model were reduced. Then, the 5 × 5 convolution was changed to two 3 × 3 convolutions, which reduced the parameters and increased the nonlinearity of the model. A 5 × 5 convolution kernel was used to replace the original convolution for extracting coarse-grained features and improving the model’s extraction ability. A multi-scale module was added to the model to enhance the multi-faceted feature extraction capability of the model. Simultaneously, the Squeeze-and-Excitation Networks (SE-Net) module was included to increase the channel’s attention and improve the model’s accuracy. The dataset’s images were rotated to reduce overfitting. For image rotation, the angle was multiplied by 30° to obtain the dataset enhanced by rotation operations of 30°, 60°, 90°, 120° and 150°. The experimental results showed that the recognition rate of this model in the behaviour of sika deer was 98.92%. Therefore, the model presented in this paper can be applied to the behaviour recognition of sika deer. The results will play an essential role in promoting animal behaviour recognition technology and animal health monitoring management.     

基于Inception与Residual组合网络的农作物病虫害识别

针对我国农作物病虫害识别方法中存在的速度慢、主观性强、所需专业知识要求高以及识别成本高等问题,提出一种基于Inception与Residual结构组合的Inception-resnet-v2网络模型的农作物病虫害识别方法,以实现精准高效的农作物病虫害识别。网络使用residual结构,采用跨层连接方式将低层特征与高层特征进行组合学习以增加网络深度。同时加入了Inception结构,既能保持网络结构的稀疏性,又能利用密集矩阵的高计算性能,加快了训练速度。最后通过Softmax分类器进行多分类预测。与传统方法相比,本文方法收敛速度更快,不仅准确率达到96.67%、精确度达到90.77%、召回率达到89.72%,还使病虫害识别的不同类别更加均衡,改善了传统方法中对特定类别识别效果差的问题。     

利用深度学习的HEVC帧内编码单元快速划分算法

新一代视频编码标准高效视频编码(High Efficiency Video Coding,HEVC)中编码单元(Coding Unit,CU)大小不同的特性使得编码效率得到显著提升,但同时带来了极高的计算复杂度.为了去除CU划分中多余的计算从而降低编码复杂度,本文提出了一种利用深度学习的编码单元快速划分算法.首先使用原始视频亮度块及编码信息建立了一个HEVC中CU划分的数据库,用于接下来本文深度学习神经网络的训练.然后,为了更好地贴合编码单元划分的层级结构,本文提出了一种基于Inception模块的神经网络结构,使之内嵌于HEVC编码框架中对编码单元的划分进行提前预测,有效地去除了All Intra配置下中冗余的CU划分计算.实验结果表明,本文提出的算法与HEVC官方测试模型(HM16.12)相比,编码时间平均降低了61.31%,而BD-BR与BD-PSNR仅为1.86%和-0.13dB.     

改进空化模型和修正湍流模型在螺旋桨空化模拟中的评估分析

为评估改进Sauer空化模型和修正切应力输运湍流模型联合使用时在螺旋桨空化模拟中的效果,以及"当σ>σi时,叶梢截面压力系数分布相对不再改变"的空化初生判定准则的适用性,以E779A螺旋桨为对象,对轻度(Ac/A0<0.1)、中度(0.1≤Ac/A0<0.25)和重度(0.25≤Ac/A0<0.5)空化程度下的桨叶空化形态、空化面积、叶截面压力系数分布、推力和力矩崩溃性能曲线以及初生空化数进行校验和分析。结果表明,在轻度空化下模拟空化面积较试验值略小;中度空化下与试验值吻合很好,模拟精度明显高于同类文献;在重度空化下因包含了泡空化面积而较试验值要大。在设计和非设计进速系数下,预报因空化引起的推力下降起始点均与试验值吻合很好,但推力衰减指数较试验值要大。应用空化初生判定准则得到J=0.77工况下初生空化数为3.25,较试验值仅偏差0.6%。由此表明,所采用的数值模型对于螺旋桨轻度和中度空化程度下的空化性能模拟具有较高的精度,所提出的空化初生判定准则是可信而且适用的。应用该准则,在求取0.88R(低侧斜桨)或0.7R(大侧斜桨)截面的初生空化数后,可较准确地预报螺旋桨空化斗图谱,进而预报舰艇空化初始航速。     

基于改进CNN的串联型故障电弧识别方法研究

针对串联型故障电弧影响供电系统安全且不易测量的问题,提出改进卷积神经网络对串联型故障电弧的识别方法.由于SVM学习的超平面是距离各个样本最远的平面,相比于Softmax,具有更强的泛化推广能力和更高的识别准确率,故采用SVM损失函数(hinge loss)替换原有的Softmax损失函数并在CNN模型中添加三层Ince...     

Probe measurements and numerical model predictions of evolving size distributions in premixed flames

Particle size distributions (PSDs), measured with a dilution probe and a Differential Mobility Analyzer (DMA), and numerical predictions of these PSDs, based on a model that includes only coagulation or alternatively inception and coagulation, are compared to investigate particle growth processes and possible sampling artifacts in the post-flame region of a C/O = 0.65 premixed laminar ethylene-air flame. Inputs to the numerical model are the PSD measured early in the flame (the initial condition for the aerosol population) and the temperature profile measured along the flame’s axial centerline. The measured PSDs are initially unimodal, with a modal mobility diameter of 2.2 nm, and become bimodal later in the post-flame region. The smaller mode is best predicted with a size-dependent coagulation model, which allows some fraction of the smallest particles to escape collisions without resulting in coalescence or coagulation through the size-dependent coagulation efficiency (γ_SD). Instead, when γ = 1 and the coagulation rate is equal to the collision rate for all particles regardless of their size, the coagulation model significantly under predicts the number concentration of both modes and over predicts the size of the largest particles in the distribution compared to the measured size distributions at various heights above the burner. The coagulation (γ_SD) model alone is unable to reproduce well the larger particle mode (mode II). Combining persistent nucleation with size-dependent coagulation brings the predicted PSDs to within experimental error of the measurements, which seems to suggest that surface growth processes are relatively insignificant in these flames. Shifting measured PSDs a few mm closer to the burner surface, generally adopted to correct for probe perturbations, does not produce a better matching between the experimental and the numerical results.     

基于改进SSD算法的车辆检测

SSD算法利用多尺度特征图进行分类和位置回归,检测小目标效果优于YOLO算法,但SSD算法在进行车辆检测时存在漏检问题。为此,提出一种改进SSD算法。为提取更多的车辆特征信息,设计改进Inception模块替代SSD网络中的Conv8、Conv9和Conv10层。将浅层特征的位置信息和深层特征的语义信息进行均衡化融合,构建多尺度特征融合均衡化网络,提高小目标车辆识别率。在特征提取层均引入SENet,对不同特征通道的重要性进行重标定以提高模型性能。实验结果表明,改进后SSD算法在自制的车辆数据集上平均精度为90.89%,检测速度达到59.42 frame/s,相比改进前的SSD算法,在精度和速度上分别提高2.65个百分点和17.41 frame/s,能够更快速、准确地对图像中的车辆进行识别和定位。     

改进的YOLOv3网络在钢板表面缺陷检测研究

为了提高工业自动化水平,对表面缺陷进行有效检测,提出了一种改进的YOLOv3（You Only Look Once）网络检测方法。使用轻量级网络（MobileNet）来代替YOLOv3原有网络中的密集连接网络（Darknet-53）,适当减少参数量的提取;加入空洞卷积,提高网络对小目标缺陷的检测能力;在网络结构的最后一层卷积中加入了Inception结构,进一步减少参数总量并加深网络。改进后的网络在测试集上精准性比原有的YOLOv3网络提高了23.3%,实时性也提高了95.4%,在钢板表面缺陷检测中具有更好的应用前景。     

基于多Inception结构的卷积神经网络人脸识别算法

人脸识别是视觉识别的一个重要领域,由于人脸识别尺度变化范围大,光照、姿态变化剧烈以及遮挡问题,导致该类非限制条件下的识别难度较大,为了解决该类问题,本文提出了一种基于Tensorflow平台的多Inception模型,通过将多个Inception结构进行串联,再通过分解卷积核的方式减少输入参数,实现了多维度同时卷积再聚合,提高了人脸识别的精度.实验结果表明,该方法在较少参数的条件下能提取出更具区分度的人脸特征,与分类损失方法及融合了其他度量学习方式的方法相比,提高了识别准确率,减少了计算时间.     

结合多层池化的卷积神经网络在表情识别中的应用

为了解决传统卷积神经网络(Convolutional Neural Network,CNN)对于人脸微表情识别的泛化能力差的问题,提出了一种改进的Inception结构与残差结构结合的卷积神经网络方法.首先在改进的Inception结构的基础上将输入特征直接映射到输出结果中构成残差结构,并针对表情局部特征复杂模糊等不足...