基于多层交叉注意力融合网络模型的人脸图像情感分析

目前大多数人脸图像情感分析方法只单方面关注图像整体或局部来构建视觉情感特征表示,忽略了二者在情感表达上的协同作用。针对此问题,提出了一种多层交叉注意力融合网络情感分析方法。该方法首先利用特征相关性分析实现最大化类的可分性;其次通过多层交叉注意力网络中的多个不重叠的注意力区域来提取整体和局部的信息;然后将整体与局部提取的注意力图进行融合,来共同训练图像情感分类器并进行情感分析。实验结果表明,提出的方法在真实数据集RAFDB上的情感分类准确率达到了88.53%,优于现有其他方法,验证了该方法的有效性与优越性。     

基于递归长短期记忆网络和镜头序列注意网络的视频摘要生成

为解决基于长短期记忆网络LSTM的视频摘要生成方法当输入序列过长时LSTM网络中的记忆单元不能集中在长时间序列的跨度上。通过深度学习的方法研究了一种基于递归长短期记忆网络(ReLSTM)和序列注意(SSA)的视频摘要生成模型用以提高深度学习网络学习时序特征的能力。该模型使用ReLSTM网络提取时间特征。同时,利用SSA动态调整每个视频序列输入到ReLSTM网络中的特征权重。结果表明：在数据集TVSum上F1-score平均提高2.5%,最高提高0.2%。在数据集SumMe上F1-score平均提高7.8%,最高提高3.4%。可见该方法能有效地学习镜头之间的时序特征。     

"文本为主"的多模态特征融合的新闻视频分类算法

视频分类算法通过融合多模态特征提高分类性能.大多数多模态特征融合算法采用统一模式处理不同模态特征,忽略文本与声音/图像特征在视频分类能力和可靠性方面的差异.该文提出一种新闻视频分类算法,采用文本为主,声音/图像为辅的特征融合模式.该算法充分考虑文本特征在内容分类方面的高可靠性,并在适当情况下,辅以声音/图像信息为补充.对10类新闻视频进行分类实验,并与其他两种分类算法对比,表明在提取相同的底层特征情况下,该文算法的平均分类性能最佳.     

Undecaborates M2[B11H11]: Facile Synthesis,Crystal Structure,and Reactions

closo-Undecaborates were synthesized by the deprotonation of B₁₁H₁₃(SMe₂) with LitBu in thp or K[BHEt₃] in thf, [Li(thp)₃]₂[B₁₁H₁₁] and K₂[B₁₁H₁₁] being obtained in 83 and 93% yield, respectively. K₂[B₁₁H₁₁] can be transformed into A₂[B₁₁H₁₁] with the corresponding ammonium chlorides in aqueous solution (A = [NMe₃Ph], [NBzlEt₃], [N(PPh₃)₂]). The crystal structure analysis of [Li(thp)₃]₂[B₁₁H₁₁] (space group P2₁/c) reveals a rather distorted octadecahedron for the [B₁₁H₁₁]^2– anion, whereas the corresponding octadecahedron in [NBzlEt₃]₂[B₁₁H₁₁] (space group P2₁2₁2₁) exhibits a structure close to C_2v symmetry, expected for the free anion. The protonation of [B₁₁H₁₁]^2– at low temperature gives [B₁₁H₁₂]^–, whose structure could be elucidated by NMR methods; it is formed, apparently, by the opening of the B1–B4 edge of [B₁₁H₁₁]^2– in the course of its known degenerate skeletal rearrangement, followed by the protonation of the B2–B4 edge. The reaction of [B₁₁H₁₂]^– with a second molecule of the acid HX (X = CF₃COO) gives nido-[B₁₁H₁₃X]^–. The addition of BH₃ to [B₁₁H₁₁]^2– yields closo-[B₁₂H₁₂]^2– under loss of H₂. Two [B₁₁H₁₁]^2– units are fused by the aid of FeCl₃, with the known anion [B₂₂H₂₂]^2– as the product, whose ¹¹B-NMR signals could completely be assigned on the basis of C_s symmetry. The compound [NBzlEt₃][N(PPh₃)₂][B₂₂H₂₂] crystallizes in the space group Pna2₁.     

基于双通道脉冲耦合神经网络的应用研究

通过对双通道脉冲神经网络模型的改进,提出了一种邻域SEML激励的自适应双通道脉冲耦合神经网络医学图像融合的新方法。即通过提高模型的性能和脉冲神经网络进行自动化图像处理的能力。首先介绍了双通道脉冲耦合神经网络模型及其原理,然后阐述了基于这一模型的一种新融合算法,最后对本文融合算法以及各传统融合算法得到的实验仿真结果进行了分析。     

一种新的彩色图像边缘检测算法

传统的边缘检测方法大都基于灰度图像,不能充分利用彩色图像的全部信息。针对已有算法中存在的像素点扩散、边缘定位不准确、边缘不连续等问题,提出了一种彩色图像边缘提取算法,基于图像自身梯度方向信息和多通道信息融合技术,将灰度边缘模板算子扩展应用到彩色图像的边缘检测中,在RGB空间中对原彩色图像进行多通道边缘检测;同时采用滤波来抑制噪声,依靠边缘生长保证检出边缘的连续性,并提出了自适应确定边缘提取门限值的方法。该文提出的彩色图像边缘检测算法计算量小,实验结果表明了其能充分利用图像的颜色和梯度信息,有效地消除噪声,提高边缘检测的准确性,保证边缘的连续性。     

加窗后波前功率谱密度的计算值修正

 波前功率谱密度的数值计算会由于窗函数的使用引入较大的计算误差。通过对模拟的单一频率波前加窗前后的功率谱密度的理论计算,由傅里叶变换性质推导出了修正因子,并对波前频率与修正因子的关系进行了理论研究。结果表明,在一定的波前频率及误差范围内,对加汉宁窗后1维功率谱密度的计算结果乘上一个常量8/3即可实现简单有效地修正。     

一种基于角点匹配的图像拼接算法研究

提出了一种基于角点匹配的图像拼接算法。研究了基于角点匹配的图像拼接方法中,Harris角点的检测,角点的匹配及图像的融合等关键算法。并且给出了图像拼接的MATLAB仿真结果。实验表明,该算法能够获得较为满意的拼接效果。     

Origin Identification of Hungarian Honey Using Melissopalynology,Physicochemical Analysis,and Near Infrared Spectroscopy

The objective of the study was to check the authenticity of Hungarian honey using physicochemical analysis, near infrared spectroscopy, and melissopalynology. In the study, 87 samples from different botanical origins such as acacia, bastard indigo, rape, sunflower, linden, honeydew, milkweed, and sweet chestnut were collected. The samples were analyzed by physicochemical methods (pH, electrical conductivity, and moisture), melissopalynology (300 pollen grains counted), and near infrared spectroscopy (NIRS:740–1700 nm). During the evaluation of the data PCA-LDA models were built for the classification of different botanical and geographical origins, using the methods separately, and in combination (low-level data fusion). PC number optimization and external validation were applied for all the models. Botanical origin classification models were >90% and >55% accurate in the case of the pollen and NIR methods. Improved results were obtained with the combination of the physicochemical, melissopalynology, and NIRS techniques, which provided >99% and >81% accuracy for botanical and geographical origin classification models, respectively. The combination of these methods could be a promising tool for origin identification of honey.     

甘蓝型油菜植物防御素基因的克隆与原核表达

根据Genbank上已知的植物防御素基因设计引物,以甘蓝型油菜中的品种”蜀杂九号”种子总RNA反转录成的eDNA为模板,进行PCR扩增,获得了243bp的片段．将该片段回收。连接到pMD18-T载体测序,将测序片段经酶切回收后克隆到GTK表达载体中,在0．1mmol IPTG诱导下表达出与理论值相符的34kD的融合蛋白条带,用GST单克隆抗体做第一抗体进行Western-blot检测,获得阳性结果,这为甘蓝型油菜植物防御素基因功能的进一步研究提供了基础．