Building indoor dangerous behavior recognition is a specific application in the field of abnormal human recognition.A human dangerous behavior recognition method based on LSTM-GCN with attention mechanism(GLA)model wa...Building indoor dangerous behavior recognition is a specific application in the field of abnormal human recognition.A human dangerous behavior recognition method based on LSTM-GCN with attention mechanism(GLA)model was proposed aiming at the problem that the existing human skeleton-based action recognition methods cannot fully extract the temporal and spatial features.The network connects GCN and LSTMnetwork in series,and inputs the skeleton sequence extracted by GCN that contains spatial information into the LSTM layer for time sequence feature extraction,which fully excavates the temporal and spatial features of the skeleton sequence.Finally,an attention layer is designed to enhance the features of key bone points,and Softmax is used to classify and identify dangerous behaviors.The dangerous behavior datasets are derived from NTU-RGB+D and Kinetics data sets.Experimental results show that the proposed method can effectively identify some dangerous behaviors in the building,and its accuracy is higher than those of other similar methods.展开更多
The healthy condition of the milling tool has a very high impact on the machining quality of the titanium components.Therefore,it is important to recognize the healthy condition of the tool and replace the damaged cut...The healthy condition of the milling tool has a very high impact on the machining quality of the titanium components.Therefore,it is important to recognize the healthy condition of the tool and replace the damaged cutter at the right time.In order to recognize the health condition of the milling cutter,a method based on the long short term memory(LSTM)was proposed to recognize tool health state in this paper.The various signals collected in the tool wear experiments were analyzed by time-domain statistics,and then the extracted data were generated by principal component analysis(PCA)method.The preprocessed data extracted by PCA is transmitted to the LSTM model for recognition.Compared with back propagation neural network(BPNN)and support vector machine(SVM),the proposed method can effectively utilize the time-domain regulation in the data to achieve higher recognition speed and accuracy.展开更多
The purpose of this paper is to solve the problem of robust face recognition(FR) with single sample per person(SSPP). In the scenario of FR with SSPP, we present a novel model local robust sparse representation(LRSR) ...The purpose of this paper is to solve the problem of robust face recognition(FR) with single sample per person(SSPP). In the scenario of FR with SSPP, we present a novel model local robust sparse representation(LRSR) to tackle the problem of query images with various intra-class variations,e.g., expressions, illuminations, and occlusion. FR with SSPP is a very difficult challenge due to lacking of information to predict the possible intra-class variation of the query images.The key idea of the proposed method is to combine a local sparse representation model and a patch-based generic variation dictionary learning model to predict the possible facial intraclass variation of the query images. The experimental results on the AR database, Extended Yale B database, CMU-PIE database and LFW database show that the proposed method is robust to intra-class variations in FR with SSPP, and outperforms the state-of-art approaches.展开更多
Digit recognition from a natural scene text in video surveillance/broadcasting applications is a challenging research task due to blurred, font variations, twisted, and non-uniform color distribution issues with a dig...Digit recognition from a natural scene text in video surveillance/broadcasting applications is a challenging research task due to blurred, font variations, twisted, and non-uniform color distribution issues with a digit in a natural scene to be recognized. In this paper, to solve the digit number recognition problem, a principal-axis based topology contour descriptor with support vector machine （SVM） classification is proposed. The contributions of this paper include： a） a local descriptor with SVM classification for digit recognition, b） higher accuracy than the state-of-the art methods, and c） low computational power （0.03 second/digit recognition）, which make this method adoptable to real-time applications.展开更多
A FPGA-based image recognition system is designed for eolorimetric sensor array in order to recognize a wide range of volatile organic compounds. The gas molecule is detected by the responsive sensor array and the res...A FPGA-based image recognition system is designed for eolorimetric sensor array in order to recognize a wide range of volatile organic compounds. The gas molecule is detected by the responsive sensor array and the responsive image is obtained. The image is decomposed to RGB color components using CMOS image sensor. An embedded image recognition archi- tecture based on Xilinx Spartan-3 FPGA is designed to implement the algorithms of image recognition. The algorithm of color coherence vector is discussed in detail compared with the algorithm of color histograms, and experimental results demonstrate that both of the two algorithms could be analyzed effectively to represent different volatile organic compounds according to their different responsive images in this system.展开更多
The housing crisis in Ireland has rapidly grown in recent years. To make a more significant profit, many landlords are no longer renting out their houses under long-term tenancies but under short-term tenancies. Regul...The housing crisis in Ireland has rapidly grown in recent years. To make a more significant profit, many landlords are no longer renting out their houses under long-term tenancies but under short-term tenancies. Regulating rentals in Rent Pressure Zones with the highest and rising rents is becoming a tricky issue. In this paper, we develop a breach identifier to check short-term rentals located in Rent Pressure Zones with potential breaches only using publicly available data from Airbnb (an online marketplace focused on short-term home-stays) and Irish government websites. First, we use a Residual Neural Network to filter out outdoor landscape photos that negatively impact identifying whether an owner has multiple rentals in a Rent Pressure Zone. Second, a Siamese Neural Network is used to compare the similarity of indoor photos to determine if multiple rental posts correspond to the same residence. Next, we use the Haversine algorithm to locate short-term rentals within a circle centered on the coordinate of a permit. Short-term rentals with a permit will not be restricted. Finally, we improve the occupancy estimation model combined with sentiment analysis, which may provide higher accuracy.展开更多
为了提高脑电情绪识别分类精度,最大限度利用脑电信号的空间和时间信息,提出一种Inception残差注意力卷积神经网络与双向长短期记忆(bi-directional long short-term memory, BiLSTM)网络相结合的新型架构时空Inception残差注意力网络...为了提高脑电情绪识别分类精度,最大限度利用脑电信号的空间和时间信息,提出一种Inception残差注意力卷积神经网络与双向长短期记忆(bi-directional long short-term memory, BiLSTM)网络相结合的新型架构时空Inception残差注意力网络。将脑电信号采集电极位置映射到二维矩阵中,采集信号作为通道,构成三维数据;将得到的三维数据输入到时空Inception残差注意力卷积网络之中,提取时空信息;将得到的特征输入到全连接层进行分类;将Inception结构引入脑电情绪识别领域,实现多尺度特征提取,并将电极映射到矩阵之中,保留电极位置信息,使用时空Inception残差注意力网络从时空两个维度获取脑电相关信息。实验表明,使用该模型对DEAP数据集进行情绪四分类可得到93.71%的准确度,相较于对比模型,识别精度提高了10%~20%。提出的模型在脑电信号情绪识别领域具有优良性能。展开更多
针对命名实体识别(NER)任务中相关模型通常仅对字符及相关词汇进行建模,未充分利用汉字特有的字形结构信息和实体类型信息的问题,提出一种融合先验知识和字形特征的命名实体识别模型。首先,采用结合高斯注意力机制的Transformer对输入...针对命名实体识别(NER)任务中相关模型通常仅对字符及相关词汇进行建模,未充分利用汉字特有的字形结构信息和实体类型信息的问题,提出一种融合先验知识和字形特征的命名实体识别模型。首先,采用结合高斯注意力机制的Transformer对输入序列进行编码,并从中文维基百科中获取实体类型的中文释义,采用双向门控循环单元(BiGRU)编码实体类型信息作为先验知识,利用注意力机制将它与字符表示进行组合;其次,采用双向长短时记忆(BiLSTM)网络编码输入序列的远距离依赖关系,通过字形编码表获得繁体的仓颉码和简体的现代五笔码,采用卷积神经网络(CNN)提取字形特征表示,并根据不同权重组合繁体与简体字形特征,利用门控机制将它与经过BiLSTM编码后的字符表示进行组合;最后,使用条件随机场(CRF)解码,得到命名实体标注序列。在偏口语化的数据集Weibo、小型数据集Boson和大型数据集PeopleDaily上的实验结果表明,与基线模型MECT(Multi-metadata Embedding based Cross-Transformer)相比,所提模型的F1值别提高了2.47、1.20和0.98个百分点,验证了模型的有效性。展开更多
文摘Building indoor dangerous behavior recognition is a specific application in the field of abnormal human recognition.A human dangerous behavior recognition method based on LSTM-GCN with attention mechanism(GLA)model was proposed aiming at the problem that the existing human skeleton-based action recognition methods cannot fully extract the temporal and spatial features.The network connects GCN and LSTMnetwork in series,and inputs the skeleton sequence extracted by GCN that contains spatial information into the LSTM layer for time sequence feature extraction,which fully excavates the temporal and spatial features of the skeleton sequence.Finally,an attention layer is designed to enhance the features of key bone points,and Softmax is used to classify and identify dangerous behaviors.The dangerous behavior datasets are derived from NTU-RGB+D and Kinetics data sets.Experimental results show that the proposed method can effectively identify some dangerous behaviors in the building,and its accuracy is higher than those of other similar methods.
基金National Natural Science Foundation of China(No.51805079)Shanghai Natural Science Foundation,China(No.17ZR1400600)Fundamental Research Funds for the Central Universities,China(No.16D110309)
文摘The healthy condition of the milling tool has a very high impact on the machining quality of the titanium components.Therefore,it is important to recognize the healthy condition of the tool and replace the damaged cutter at the right time.In order to recognize the health condition of the milling cutter,a method based on the long short term memory(LSTM)was proposed to recognize tool health state in this paper.The various signals collected in the tool wear experiments were analyzed by time-domain statistics,and then the extracted data were generated by principal component analysis(PCA)method.The preprocessed data extracted by PCA is transmitted to the LSTM model for recognition.Compared with back propagation neural network(BPNN)and support vector machine(SVM),the proposed method can effectively utilize the time-domain regulation in the data to achieve higher recognition speed and accuracy.
基金supported in part by the National Natural Science Foundation of China(61673402,61273270,60802069)the Natural Science Foundation of Guangdong Province(2017A030311029,2016B010109002,2015B090912001,2016B010123005,2017B090909005)+1 种基金the Science and Technology Program of Guangzhou of China(201704020180,201604020024)the Fundamental Research Funds for the Central Universities of China
文摘The purpose of this paper is to solve the problem of robust face recognition(FR) with single sample per person(SSPP). In the scenario of FR with SSPP, we present a novel model local robust sparse representation(LRSR) to tackle the problem of query images with various intra-class variations,e.g., expressions, illuminations, and occlusion. FR with SSPP is a very difficult challenge due to lacking of information to predict the possible intra-class variation of the query images.The key idea of the proposed method is to combine a local sparse representation model and a patch-based generic variation dictionary learning model to predict the possible facial intraclass variation of the query images. The experimental results on the AR database, Extended Yale B database, CMU-PIE database and LFW database show that the proposed method is robust to intra-class variations in FR with SSPP, and outperforms the state-of-art approaches.
基金supported by“MOST”under Grant No.105-2221-E-119-001
文摘Digit recognition from a natural scene text in video surveillance/broadcasting applications is a challenging research task due to blurred, font variations, twisted, and non-uniform color distribution issues with a digit in a natural scene to be recognized. In this paper, to solve the digit number recognition problem, a principal-axis based topology contour descriptor with support vector machine （SVM） classification is proposed. The contributions of this paper include： a） a local descriptor with SVM classification for digit recognition, b） higher accuracy than the state-of-the art methods, and c） low computational power （0.03 second/digit recognition）, which make this method adoptable to real-time applications.
基金supported by the National Natural Science Foundation of China under Grant No.60771044
文摘A FPGA-based image recognition system is designed for eolorimetric sensor array in order to recognize a wide range of volatile organic compounds. The gas molecule is detected by the responsive sensor array and the responsive image is obtained. The image is decomposed to RGB color components using CMOS image sensor. An embedded image recognition archi- tecture based on Xilinx Spartan-3 FPGA is designed to implement the algorithms of image recognition. The algorithm of color coherence vector is discussed in detail compared with the algorithm of color histograms, and experimental results demonstrate that both of the two algorithms could be analyzed effectively to represent different volatile organic compounds according to their different responsive images in this system.
文摘The housing crisis in Ireland has rapidly grown in recent years. To make a more significant profit, many landlords are no longer renting out their houses under long-term tenancies but under short-term tenancies. Regulating rentals in Rent Pressure Zones with the highest and rising rents is becoming a tricky issue. In this paper, we develop a breach identifier to check short-term rentals located in Rent Pressure Zones with potential breaches only using publicly available data from Airbnb (an online marketplace focused on short-term home-stays) and Irish government websites. First, we use a Residual Neural Network to filter out outdoor landscape photos that negatively impact identifying whether an owner has multiple rentals in a Rent Pressure Zone. Second, a Siamese Neural Network is used to compare the similarity of indoor photos to determine if multiple rental posts correspond to the same residence. Next, we use the Haversine algorithm to locate short-term rentals within a circle centered on the coordinate of a permit. Short-term rentals with a permit will not be restricted. Finally, we improve the occupancy estimation model combined with sentiment analysis, which may provide higher accuracy.
文摘为了提高脑电情绪识别分类精度,最大限度利用脑电信号的空间和时间信息,提出一种Inception残差注意力卷积神经网络与双向长短期记忆(bi-directional long short-term memory, BiLSTM)网络相结合的新型架构时空Inception残差注意力网络。将脑电信号采集电极位置映射到二维矩阵中,采集信号作为通道,构成三维数据;将得到的三维数据输入到时空Inception残差注意力卷积网络之中,提取时空信息;将得到的特征输入到全连接层进行分类;将Inception结构引入脑电情绪识别领域,实现多尺度特征提取,并将电极映射到矩阵之中,保留电极位置信息,使用时空Inception残差注意力网络从时空两个维度获取脑电相关信息。实验表明,使用该模型对DEAP数据集进行情绪四分类可得到93.71%的准确度,相较于对比模型,识别精度提高了10%~20%。提出的模型在脑电信号情绪识别领域具有优良性能。
文摘针对命名实体识别(NER)任务中相关模型通常仅对字符及相关词汇进行建模,未充分利用汉字特有的字形结构信息和实体类型信息的问题,提出一种融合先验知识和字形特征的命名实体识别模型。首先,采用结合高斯注意力机制的Transformer对输入序列进行编码,并从中文维基百科中获取实体类型的中文释义,采用双向门控循环单元(BiGRU)编码实体类型信息作为先验知识,利用注意力机制将它与字符表示进行组合;其次,采用双向长短时记忆(BiLSTM)网络编码输入序列的远距离依赖关系,通过字形编码表获得繁体的仓颉码和简体的现代五笔码,采用卷积神经网络(CNN)提取字形特征表示,并根据不同权重组合繁体与简体字形特征,利用门控机制将它与经过BiLSTM编码后的字符表示进行组合;最后,使用条件随机场(CRF)解码,得到命名实体标注序列。在偏口语化的数据集Weibo、小型数据集Boson和大型数据集PeopleDaily上的实验结果表明,与基线模型MECT(Multi-metadata Embedding based Cross-Transformer)相比,所提模型的F1值别提高了2.47、1.20和0.98个百分点,验证了模型的有效性。
