A Smart Deep Convolutional Neural Network for Real-Time Surface Inspection

A proper detection and classification of defects in steel sheets in real time have become a requirement for manufacturing these products, largely used in many industrial sectors. However, computers used in the production line of small to medium size companies, in general, lack performance to attend real-time inspection with high processing demands. In this paper, a smart deep convolutional neural network for using in real-time surface inspection of steel rolling sheets is proposed. The architecture is based on the state-of-the-art SqueezeNet approach, which was originally developed for usage with autonomous vehicles. The main features of the proposed model are: small size and low computational burden. The model is 10 to 20 times smaller when compared to other networks designed for the same task, and more than 700 times smaller than general networks. Also, the number of floating-point operations for a prediction is about 50 times lower than the ones used for similar tasks. Despite its small size, the proposed model achieved near-perfect accuracy on a public dataset of 1800 images of six types of steel rolling defects.     

Sequential alignment attention model for scene text recognition

Scene text recognition has been a hot research topic in computer vision due to its various applications. The state-of-the-art solutions usually depend on the attention-based encoder-decoder framework that learns the mapping between input images and output sequences in a purely data-driven way. Unfortunately, there often exists severe misalignment between feature areas and text labels in real-world scenarios. To address this problem, this paper proposes a sequential alignment attention model to enhance the alignment between input images and output character sequences. In this model, an attention gated recurrent unit (AGRU) is first devised to distinguish the text and background regions, and further extract the localized features focusing on sequential text regions. Furthermore, CTC guided decoding strategy is integrated into the popular attention-based decoder, which not only helps to boost the convergence of the training but also enhances the well-aligned sequence recognition. Extensive experiments on various benchmarks, including the IIIT5k, SVT, and ICDAR datasets, show that our method substantially outperforms the state-of-the-art methods.     

投影寻踪分类模型在常见造纸纤维原料综合评价中的应用

将投影寻踪分类（PPC）模型与基于实数编码的加速遗传算法（RAGA）相结合，同时优化多个指标参数，将高维数据指标转化为低维空间上的一维投影值，建立RAGA-PPC模型，用于造纸纤维原料分类，并对造纸纤维原料进行综合评价。结果表明，基于RAGA-PPC模型的评价结果与造纸纤维原料实际分类结果一致，此方法客观可靠，精度高，具有一定的应用前景。     

基于全相位预处理的时域多重信号分类波达方向估计方法

低信噪比下,针对宽带短脉冲情况下频域多重信号分类(MUSIC)中噪声子空间估计不稳定问题,提出一种基于全相位预处理的时域多重信号分类波达方向(DOA)估计方法。①对线列阵接收数据进行分组处理;②按搜索角度对各组数据进行相移预处理,并对各组数据预处理结果进行相加,得到一组新数据;③对线列阵接收数据在时域构建相移后的协方差矩阵,在更短数据长度下,稳定实现噪声子空间估计,并依据估计出的噪声子空间含有的正交特性,通过单位矩阵加法器得到相应空间谱估计值,实现波达方向估计。数值仿真和实测数据处理结果表明,相比频域MUSIC方法,该方法有效提高了线列阵接收数据协方差矩阵中信号含有量和信噪比,能够在更短数据长度情况下实现对噪声子空间的稳定估计,具有较好的稳定性和检测性能,提高了MUSIC方法在实际波达方向估计中的鲁棒性。     

带权决策表的属性约简

属性约简是粗糙集理论的重要应用。考虑将决策表中的每行都作为一条决策规则时，若把表中出现相同决策规则的次数作为权，可得到带权决策表。提出了关于带权决策表的正域约简相应的辨识矩阵并给出了证明，从而得到了约简算法。相比于决策表中的正域约简时发现，通过将决策表转化为带权决策表后，再利用算法1进行约简时，其在一定程度上优于前者。提出了近似分类精度约简相应的辨识矩阵并给出了证明。对于2个算法，在选取的UCI数据集上进行了实验验证。通过实验进一步说明了所提出算法的可行性和有效性。     

基于改进的Random Subspace的客户投诉分类方法

电信业的客户投诉不断增多而又亟待高效处理。针对电信客户投诉数据的特点，提出了一种面向高维数据的改进的集成学习分类方法。该方法综合考虑客户投诉中的文本信息及客户通讯状态信息，基于Random Subspace方法，以支持向量机（Support Vector Machine，SVM）为基分类器，采用证据推理（Evidential Reasoning，ER）规则为一种新的集成策略，构造分类模型对电信客户投诉进行分类。所提模型和方法在某电信公司客户投诉数据上进行了验证，实验结果显示该方法能够显著提高客户投诉分类的准确率和投诉处理效率。     

结合预训练模型和语言知识库的文本匹配方法

针对文本匹配任务,该文提出一种大规模预训练模型融合外部语言知识库的方法。该方法在大规模预训练模型的基础上,通过生成基于WordNet的同义—反义词汇知识学习任务和词组—搭配知识学习任务引入外部语言学知识。进而,与MT-DNN多任务学习模型进行联合训练,以进一步提高模型性能。最后利用文本匹配标注数据进行微调。在MRPC和QQP两个公开数据集的实验结果显示,该方法可以在大规模预训练模型和微调的框架基础上,通过引入外部语言知识进行联合训练有效提升文本匹配性能。     

面向智能客服系统的情感分析技术

该文以阿里小蜜为例,对智能客服系统中的情感分析技术进行比较全面的介绍,包括情感分析算法模型的原理及其在智能客服系统的多个应用场景中的实际落地使用方式和效果分析。智能客服在解决客户高频业务问题的同时,也需要给客户提供多维度的、具有类人能力的助理、导购、语聊和娱乐等服务,提高客户对智能客服机器人的整体满意度。在此过程中,情感分析技术在机器人类人能力建设中起到了至关重要的作用。该文围绕智能客服系统中人机结合的服务形式,从六个维度总结和介绍了情感分析技术在智能客服系统中的应用场景,包括用户情感检测、用户情感安抚、情感生成式语聊、客服服务质检、会话满意度预估和智能人工入口。     

Development of a statistical model to classify driving stress levels using galvanic skin responses

Many studies have demonstrated the strong relationships between physiological responses and driving stress, but they have done little to build a model that could be used to identify a driver's stress accurately in real time. The objective of this study is to develop a model that accurately classifies driving stress by monitoring physiological responses—specifically galvanic skin response (GSR). GSR data were collected from nine drivers with licenses obtained in the US in real road driving situations with two stress conditions—rest period (low stress) and highway or city driving (high stress). The validation drive was performed by one driver with licenses obtained in South Korea in real long‐term road driving situations with two stress conditions—rural area (low stress) and highway or highway under construction (high stress). Those two conditions were used to build a binary logistic regression model to classify low stress or high stress based on a driver's measured hand GSR. The overall classification accuracy of the developed model was found to be 85.3%, and the accuracy of cross validation, with a testing dataset, was found to be 83.2%. A simple logit model was developed to identify drivers' stress by incorporating their GSR data. The developed model can be embedded in a wearable device equipped with GSR sensors for drivers to detect their stress level in real time.