结合注意力与局部特征融合的行人重识别算法

行人重识别是指从一堆候选图片中找到与目标最相似的行人图片,本质上是一个图像检索的子问题。为了进一步增强网络提取关键特征的能力以及抑制噪声的干扰,通过对基于注意力机制和局部特征的行人重识别算法的研究,提出了结合注意力与局部特征融合的行人重识别算法。该算法将ResNeSt-50作为骨干网络,联合软注意力与非局部注意力机制,采用双流结构分别提取行人细粒度全局特征和细粒度局部特征,通过关注不同特征之间共享的空间域信息以及同一特征不同水平区域的潜在语义相关性,创建了空间感知特征融合模块（spatial-aware feature fusion module）以及跨区域特征融合模块（cross-region feature fusion module）。在Market-1501、DukeMTMC-reID以及CUHK03数据集上的实验结果表明该算法极大程度上提升了网络的检索能力,同时与现有算法进行比较,凸显出优越性能。     

基于局部特征注意力的行人重识别方法

行人重识别旨在多个视频传感器条件下,从图像库中出检索特定的行人目标,具有重要的实际应用价值。针对以往对局部特征利用不足的情况,创新一种基于注意力引导的局部特征关系融合方法,使在对局部特征分别计算的同时,通过注意力引导,探索各局部特征之间的内部关系。首先将图像通过残差网络ResNet-50获取特征,然后对特征进行水平分割获取局部特征后,通过注意力引导的局部特征关系融合网络,最后使用难采样三元组损失函数和交叉熵损失函数对模型进行训练。实验表明,该算法在行人重识别公开数据集Market-1501上mAP值达到86.4%,Rank-1达到94.7%。     

用于车辆重识别的部件耦合Transformer网络

基于卷积神经网络的车辆重识别模型在执行卷积和池化操作时,不可避免地会出现全局感受野狭小和局部信息丢失的情况,当光照、视角和分辨率等发生剧烈变化时,导致车辆重识别的鲁棒性和精确性急剧下降.为此,提出了部件耦合Transformer的车辆重识别网络,通过堆叠部件耦合Transformer块来搭建重识别模型,每一个部件耦合Transformer块利用部件自适应嵌入模块提取区分性的局部特征和Transformer层提取鲁棒性的全局特征.首先,部件自适应嵌入模块按照位置和伸缩量动态划分和调整特征图,增强模型对局部部件信息的感知能力;其次, Transformer层中利用自注意力机制增强网络模型对全局特征的表示能力;最后,部件自适应嵌入模块和Transformer层之间的耦合关系促进全局和局部特征协同合作.在VeRi-776和VehicleID数据集上的实验结果表明,CMC@1/CMC@5分别达到0.970/0.988和0.865/0.985,优于对比模型.     

联合语义分割与注意力机制的行人再识别模型

受行人姿态变化、光照视角、背景变换等因素的影响,现有行人再识别模型通常对数据集中的行人分成若干块提取图像的局部特征进行辨识以提高识别精度,但存在人体局部特征不匹配、容易丢失非人体部件的上下文线索等问题。构建一种改进的行人再识别模型,通过将人体语义解析网络的局部特征进行对齐,增强行人语义分割模型对图像中行人任意轮廓的建模能力,利用局部注意力网络捕捉非人体部分丢失的语境线索。实验结果表明,该模型在Market-1501、DukeMTMC和CUHK03数据集上的平均精度均值分别达到83.5%、80.8%和92.4%,在DukeMTMC数据集上的Rank-1为90.2%,相比基于注意力机制、行人语义解析和局部对齐网络的行人再识别模型具有更强的鲁棒性和迁移性。     

结合注意力与批特征擦除的行人重识别模型

在行人重识别过程中,图像局部遮挡会造成识别准确率下降。提出一种结合注意力和批特征擦除的网络（ABFE-Net）模型,旨在学习具有辨别力的全局特征和局部细粒度特征,提高图像局部遮挡条件下行人特征的表达能力。将轻量级注意力模块嵌入到ResNet-50中自主学习每个通道的权重,通过强化有用特征和抑制无关特征增强网络特征的学习能力,提取行人更具辨别力的全局特征。对于深层特征使用批特征擦除方法,随机擦除同一批次特征图的相同区域,使得网络关注剩余的局部细粒度特征。将两种特征融合得到更加全面的行人特征表示,对其进行相似性度量并排序,得到行人重识别的结果。实验结果表明,与HA-CNN、PCB等方法相比,ABFE-Net模型在Market1501和DukeMTMC-reID数据集上的Rank-1和mAP分别达到94.4%、85.9%和88.3%、75.1%,能够明显增强行人特征的辨别性,提高行人重识别效果。     

融合关键点属性与注意力表征的人脸表情识别

人脸的表情变化非常细微,通常表现在图像中某些局部点区域的改变,现有的人脸表情识别方法难以捕捉到表情的细微变化,对非表情区域干扰不具有鲁棒性。为了获得描述人脸表情变化的高效特征表示,提出了一种融合关键点属性与注意力表征的人脸表情识别方法。通过添加通道注意力和空间注意力的神经网络提取人脸图像中的关键点信息,实现不同维度和位置的权重分配,有效避免非表情区域的干扰,捕获图像中局部关键点的特征表征。引入Transformer模块学习不同关键点之间的相关联系,引导网络构建对表情类型更具分辨力的特征表示,从而实现精准识别。通过在CK+、JAFFE、FER2013三种公开数据集上进行实验的结果表明：提出算法的识别准确率分别达到了99.22%、96.57%、73.37%。     

基于Transformer编码器的中文命名实体识别模型

命名实体识别是自然语言处理中的重要任务,且中文命名实体识别相比于英文命名实体识别任务更具难度。传统中文实体识别模型通常基于深度神经网络对文本中的所有字符打上标签,再根据标签序列识别命名实体,但此类基于字符的序列标注方式难以获取词语信息。提出一种基于Transformer编码器的中文命名实体识别模型,在字嵌入过程中使用结合词典的字向量编码方法使字向量包含词语信息,同时针对Transformer编码器在注意力运算时丢失字符相对位置信息的问题,改进Transformer编码器的注意力运算并引入相对位置编码方法,最终通过条件随机场模型获取最优标签序列。实验结果表明,该模型在Resume和Weibo中文命名实体识别数据集上的F1值分别达到94.7%和58.2%,相比于基于双向长短期记忆网络和ID-CNN的命名实体识别模型均有所提升,具有更优的识别效果和更快的收敛速度。     

基于改进Transformer模型的多声源分离方法

目前主流的语音分离算法模型都是基于复杂的递归网络或Transformer网络,Transformer网络复杂度高导致训练难度大以及音频的高采样率导致在样本级别上使用超长输入从而获取不完全特征,不能直接对长语音特征序列进行直接建模出现特征丢失问题。对此,该文提出了一种基于Transformer的改进网络模型。首先,在原有Transformer网络模型编码器里新添加下采样块,计算不同时间尺度上的高级特征同时降低特征空间复杂度;其次,在Transformer网络模型的解码器里添加上采样层与编码器下采样层特征融合保证特征不丢失,提高模型分离能力;最后,在模型分离层里引入一种改进的滑动窗口注意力机制,滑动窗口使用循环移位技术,新的特征窗口中包含老的特征窗口特征同时融合特征边缘信息完成了特征窗口之间的信息交互,获得特征编码以及特征位置编码同时提高特征信息之间的相关系数。实验表明,使用SI-SNR评价标准达到13.5 dB,使用SDR评价指标达到14.1 dB,分离效果优于之前的方法。     

基于注意力机制特征融合的中文命名实体识别

命名实体识别是自然语言处理领域中信息抽取、信息检索、知识图谱等任务的基础。在命名实体识别任务中,Transformer编码器更加关注全局语义,对位置和方向信息不敏感,而双向长短期记忆（BiLSTM）网络可以提取文本中的方向信息,但缺少全局语义信息。为同时获得全局语义信息和方向信息,提出使用注意力机制动态融合Transformer编码器和BiLSTM的模型。使用相对位置编码和修改注意力计算公式对Transformer编码器进行改进,利用改进的Transformer编码器提取全局语义信息,并采用BiLSTM捕获方向信息。结合注意力机制动态调整权重,深度融合全局语义信息和方向信息以获得更丰富的上下文特征。使用条件随机场进行解码,实现实体标注序列预测。此外,针对Word2Vec等传统词向量方法无法表示词的多义性问题,使用RoBERTa-wwm预训练模型作为模型的嵌入层提供字符级嵌入,获得更多的上下文语义信息和词汇信息,增强实体识别效果。实验结果表明,该方法在中文命名实体识别数据集Resume和Weibo上F1值分别达到96.68%和71.29%,相比ID-CNN、BiLSTM、CAN-NER等...     

融合先验信息的残差空间注意力人脸超分辨率重建模型

针对公共场所监控图像中低分辨率人脸图像利用现有人脸识别系统识别准确率低的问题,提出了融合先验信息的残差空间注意力人脸超分辨率重建模型,用该模型对低分辨率人脸图像进行预处理后再进行识别可大大提升识别准确率.该模型将面部先验结构信息嵌入到生成对抗网络模型中,再采用残差空间注意力激活算法突出空间位置中携带高频信息的特征,最后使用多阶特征融合算法充分利用不同尺度的特征,防止携带高频信息的人脸特征在网络传播中丢失.实验结果表明,重建出的超分辨率人脸图像具有更多的面部细节特征,大大提高了对低分辨率人脸图像的识别准确率,并且与其他5种模型相比,新模型具有较低的耗时和较少的参数.     

European Community policy and the market

Abstract This paper starts with some reflections on the policy considerations and priorities which are shaping European Commission (EC) research programmes. Then it attempts to position the current projects which seek to capitalise on information and communications technologies for learning in relation to these priorities and the apparent realities of the marketplace. It concludes that while there are grounds to be optimistic about the contribution EC programmes can make to the efficiency and standard of education and training, they are still too technology driven.     

一种自适应子融合集成多分类器方法

融合集成方法已经广泛应用在模式识别领域,然而一些基分类器实时性能稳定性较差,导致多分类器融合性能差,针对上述问题本文提出了一种新的基于多分类器的子融合集成分类器系统。该方法考虑在度量层融合层次之上通过对各类基多分类器进行动态选择,票数最多的类别作为融合系统中对特征向量识别的类别,构成一种新的自适应子融合集成分类器方法。实验表明,该方法比传统的分类器以及分类融合方法识别准确率明显更高,具有更好的鲁棒性。     

Explanation and prediction: an architecture for default and abductive reasoning

Although there are many arguments that logic is an appropriate tool for artificial intelligence, there has been a perceived problem with the monotonicity of classical logic. This paper elaborates on the idea that reasoning should be viewed as theory formation where logic tells us the consequences of our assumptions. The two activities of predicting what is expected to be true and explaining observations are considered in a simple theory formation framework. Properties of each activity are discussed, along with a number of proposals as to what should be predicted or accepted as reasonable explanations. An architecture is proposed to combine explanation and prediction into one coherent framework. Algorithms used to implement the system as well as examples from a running implementation are given.     

Three Process Perspectives: Organizations, Teams, and People

This paper provides the author's personal views and perspectives on software process improvement. Starting with his first work on technology assessment in IBM over 20 years ago, Watts Humphrey describes the process improvement work he has been directly involved in. This includes the development of the early process assessment methods, the original design of the CMM, and the introduction of the Personal Software Process (PSP)^SM and Team Software Process (TSP){^SM}. In addition to describing the original motivation for this work, the author also reviews many of the problems he and his associates encountered and why they solved them the way they did. He also comments on the outstanding issues and likely directions for future work. Finally, this work has built on the experiences and contributions of many people. Mr. Humphrey only describes work that he was personally involved in and he names many of the key contributors. However, so many people have been involved in this work that a full list of the important participants would be impractical.     

基于复小波噪声方差显著修正的SAR图像去噪

提出了一种基于复小波域统计建模与噪声方差估计显著性修正相结合的合成孔径雷达(Synthetic Aperture Radar,SAR)图像斑点噪声滤波方法。该方法首先通过对数变换将乘性噪声模型转化为加性噪声模型,然后对变换后的图像进行双树复小波变换(Dualtree Complex Wavelet Transform,DCWT),并对复数小波系数的统计分布进行建模。在此先验分布的基础上,通过运用贝叶斯估计方法从含噪系数中恢复原始系数,达到滤除噪声的目的。实验结果表明该方法在去除噪声的同时保留了图像的细节信息,取得了很好的降噪效果。     

How do children do mathematics with LOGO?

Abstract  This paper considers some results of a study designed to investigate the kinds of mathematical activity undertaken by children (aged between 8 and 11) as they learned to program in LOGO. A model of learning modes is proposed, which attempts to describe the ways in which children used and acquired understanding of the programming/mathematical concepts involved. The remainder of the paper is concerned with discussing the validity and limitations of the model, and its implications for further research and curriculum development.     

Editorial: Special issue on recent advances in hybrid control theory and applications

正The demands of a rapidly advancing technology for faster and more accurate controllers have always had a strong influence on the progress of automatic control theory.In recent years control problems have been arising with increasing frequency in widely different areas,which cannot be addressed using conventional control techniques.The principal reason for this is the fact that a highly competitive economy is forcing systems to operate in regimes where     

Information for Authors

正Aim The Journals of Zhejiang University-SCIENCE(A/B/C)areedited by the international board of distinguished Chinese andforeign scientists,and are aimed to present the latest devel-opments and achievements in scientific research in China andoverseas to the world’s scientific circles,especially to stimulateand promote academic exchange between Chinese and for-eign scientists everywhere.     

Enhancing a leaf radiative transfer model to estimate concentrations and in vivo specific absorption coefficients of total carotenoids and chlorophylls a and b from single-needle reflectance and transmittance

The relative concentrations of different pigments within a leaf have significant physiological and spectral consequences. Photosynthesis, light use efficiency, mass and energy exchange, and stress response are dependent on relationships among an ensemble of pigments. This ensemble also determines the visible characteristics of a leaf, which can be measured remotely and used to quantify leaf biochemistry and structure. But current remote sensing approaches are limited in their ability to resolve individual pigments. This paper focuses on the incorporation of three pigments—chlorophyll a, chlorophyll b, and total carotenoids—into the LIBERTY leaf radiative transfer model to better understand relationships between leaf biochemical, biophysical, and spectral properties.Pinus ponderosa and Pinus jeffreyi needles were collected from three sites in the California Sierra Nevada. Hemispheric single-leaf visible reflectance and transmittance and concentrations of chlorophylls a and b and total carotenoids of fresh needles were measured. These data were input to the enhanced LIBERTY model to estimate optical and biochemical properties of pine needles. The enhanced model successfully estimated reflectance (RMSE = 0.0255, BIAS = 0.00477, RMS%E = 16.7%), had variable success estimating transmittance (RMSE = 0.0442, BIAS = 0.0294, RMS%E = 181%), and generated very good estimates of carotenoid concentrations (RMSE = 2.48 µg/cm², BIAS = 0.143 µg/cm², RMS%E = 20.4%), good estimates of chlorophyll a concentrations (RMSE = 10.7 µg/cm², BIAS = − 0.992 µg/cm², RMS%E = 21.1%), and fair estimates of chlorophyll b concentrations (RMSE = 7.49 µg/cm², BIAS = − 2.12 µg/cm², RMS%E = 43.7%). Overall root mean squared errors of reflectance, transmittance, and pigment concentration estimates were lower for the three-pigment model than for the single-pigment model. The algorithm to estimate three in vivo specific absorption coefficients is robust, although estimated values are distorted by inconsistencies in model biophysics. The capacity to invert the model from single-leaf reflectance and transmittance was added to the model so it could be coupled with vegetation canopy models to estimate canopy biochemistry from remotely sensed data.     

Mr. Twinn’s bombes

This article discusses the history and design of the special versions of the bombe key-finding machines used by Britain’s Government Code & Cypher School (GC&CS) during World War II to attack the Enigma traffic of the Abwehr (the German military intelligence service). These special bombes were based on the design of their more numerous counterparts used against the traffic of the German armed services, but differed from them in important ways that highlight the adaptability of the British bombe design, and the power and flexibility of the diagonal board. Also discussed are the changes in the Abwehr indicating system that drove the development of these machines, the ingenious ways in which they were used, and some related developments involving the bombes used by the U.S. Navy’s cryptanalytic unit (OP-20-G).