一种弱纹理目标立体匹配网络

鉴于传统深度估计方法在高分辨率图像下存在特征提取不够充分、图像信息获取不完整、受限于局部信息或特定类型的特征提取等问题,为此提出一种面向全局特征的Transformer立体匹配网络。该网络采用编码器-解码器的端到端架构,使用多头注意力机制,允许模型在不同子空间中关注不同的特征,从而提高建模能力。模型将自注意力机制和特征重构窗口相结合,能够提高特征的表征能力,弥补局部特征不足问题,减少计算负担的同时有效应对Transformer架构通常伴随的高计算复杂度问题,确保模型的注意力计算保持在线性复杂度范围内。在Scene Flow、KITTI-2015数据集上分别进行实验,指标获得显著提升,通过对比实验验证模型的有效性和正确性。     

A Survey of Text Summarization Approaches Based on Deep Learning

Automatic text summarization (ATS) has achieved impressive performance thanks to recent advances in deep learning (DL) and the availability of large-scale corpora.The key points in ATS are to estimate the salience of information and to generate coherent results.Recently,a variety of DL-based approaches have been developed for better considering these two aspects.However,there is still a lack of comprehensive literature review for DL-based ATS approaches.The aim of this paper is to comprehensively review significant DL-based approaches that have been proposed in the literature with respect to the notion of generic ATS tasks and provide a walk-through of their evolution.We first give an overview of ATS and DL.The comparisons of the datasets are also given,which are commonly used for model training,validation,and evaluation.Then we summarize single-document summarization approaches.After that,an overview of multi-document summarization approaches is given.We further analyze the performance of the popular ATS models on common datasets.Various popular approaches can be employed for different ATS tasks.Finally,we propose potential research directions in this fast-growing field.We hope this exploration can provide new insights into future research of DL-based ATS.     

光纤扩频通信系统中光编解码器的研究

光编解码器是光纤扩频通信系统中的关键技术。本文分析了光编解码器的工作原理及其最佳设计参数的选取，采用实际制作的光编解码器进行了系统实验。获得与理论分析一致的结果。     

Anomaly detection in smart grid based on encoder-decoder framework with recurrent neural network

Anomaly detection in smart grid is critical to enhance the reliability of power systems. Excessive manpower has to be involved in analyzing the measurement data collected from intelligent motoring devices while performance of anomaly detection is still not satisfactory. This is mainly because the inherent spatio-temporality and multi-dimensionality of the measurement data cannot be easily captured. In this paper, we propose an anomaly detection model based on encoder-decoder framework with recurrent neural network (RNN). In the model, an input time series is reconstructed and an anomaly can be detected by an unexpected high reconstruction error. Both Manhattan distance and the edit distance are used to evaluate the difference between an input time series and its reconstructed one. Finally, we validate the proposed model by using power demand data from University of California, Riverside (UCR) time series classification archive and IEEE 39 bus system simulation data. Results from the analysis demonstrate that the proposed encoder-decoder framework is able to successfully capture anomalies with a precision higher than 95%.     

一种基于改进的MobileNetV2网络语义分割算法

基于金字塔卷积神经网络的语义分割算法准确率很高,但是其计算资源消耗巨大、算法执行时间长、无法满足实时性要求.为了解决这个问题,本文做出了以下改进：（1）用MobileNet替换原网络的结构,减少了网络运算时间和内存开销;（2）引入编码器-解码器结构提高输出图像的分辨率,进一步细化分割结果;（3）针对高分辨率图像推断时间过长的问题,本文设计了多级图像输入方法,降低了网络推断高分辨率图像所消耗的时间.本文在VOC 2012数据集和Cityscapes数据集上进行了测试,并与FCN、SegNet、DeepLab、PSPNet以及DFN等语义分割模型对比.实验结果表明,本文设计的语义分割算法在VOC 2012数据集上达到了76.1%的mIoU,在Cityscapes数据集上达到了74.1%的mIoU,略低于传统语义分割算法;处理一张分辨率为1024×512的图片需要18ms,少于传统语义分割算法,满足了实时性要求,达到了准确率与计算资源消耗之间的平衡.     

MU-GAN: Facial Attribute Editing Based on Multi-Attention Mechanism

Facial attribute editing has mainly two objectives: 1) translating image from a source domain to a target one, and 2) only changing the facial regions related to a target attribute and preserving the attribute-excluding details. In this work, we propose a multi-attention U-Net-based generative adversarial network (MU-GAN). First, we replace a classic convolutional encoder-decoder with a symmetric U-Net-like structure in a generator, and then apply an additive attention mechanism to build attention-based U-Net connections for adaptively transferring encoder representations to complement a decoder with attribute-excluding detail and enhance attribute editing ability. Second, a self-attention (SA) mechanism is incorporated into convolutional layers for modeling long-range and multi-level dependencies across image regions. Experimental results indicate that our method is capable of balancing attribute editing ability and details preservation ability, and can decouple the correlation among attributes. It outperforms the state-of-the-art methods in terms of attribute manipulation accuracy and image quality. Our code is available at https://github.com/SuSir1996/MU-GAN.      

采用多输入浮栅MOS器件的四值编-译码电路设计

提出一种采用多输入浮栅MOS管设计具有可控阈值功能的电压型多值逻辑电路的方法.对每个浮栅MOS管的逻辑功能均采用传输开关运算予以表示以实现有效综合。在此基础上提出了一种新的电压型多输入浮栅MOS四值编码器和译码器设计。所提出的电路在结构上得到了非常明显的简化,并可采用标准的双层多晶硅CMOS工艺予以实现。此外,这些电路具有逻辑摆幅完整、延迟小等特点。采用TSMC0.35μm双层多晶硅CMOS工艺参数的HSPICE模拟结果验证了所提出设计方案的正确性。     

基于交叉注意力机制的波束形成后置滤波网络

针对经典后置滤波器存在的对非平稳噪声抑制效果较差且存在目标语音失真的问题,提出一种基于交叉注意力机制的后置滤波网络,使用基于门控循环单元的编解码器作为网络框架,并在编解码器组之间添加残差连接;使用基于伽马通域的波束输出信号与噪声参考信号功率谱的子带增益作为双特征输入;使用特征交叉的多头归一化点积注意力捕获序列输入的长距离依赖信息并进行特征融合。实验结果表明,该算法在不同信扰比和噪声条件下的语音质量和可懂度指标均优于基线系统,具有较强的鲁棒性;在对非平稳噪声具有较好抑制效果的同时,能最小化目标语音的失真;且相较端到端的深度学习方法,具有轻量化和低时延的特点,能满足实际工程应用的需求。     

基于U-Net融合的保留纹理的图像去噪方法

图像的噪声阻碍了高级视觉任务对图像的理解,且去除图像的噪声是一个具有挑战性的任务.现有的基于卷积神经网络的图像去噪方法在去除噪声的同时,对图像纹理会引入一定程度的破坏,导致去噪后图像无法保留图像的纹理.为了解决这个问题,本文提出一种用二分支U-Net网络来融合特征和保留纹理的图像去噪方法.首先选取一种去噪方法的两个不同去噪参数的预训练模型分别得到同一张噪声图像的不同去噪结果,其中一个结果中去噪效果比纹理保留效果好,另一个结果中纹理保留比去噪效果好.然后将这两个去噪图像作为卷积神经网络的输入,利用两个编码器分别提取图像的特征,并同时放入融合模块融合图像的特征,最后利用解码器重建出无噪声图像.实验结果表明,与现有的方法相比本文的方法更有效,在去除噪声的同时能保留更多的图像纹理信息.     

基于门控化上下文感知网络的词语释义生成方法

传统的词典编纂工作主要采用人工编纂的方式,效率较低且耗费大量的资源。为减少人工编纂的时间和经济成本,该文提出一种基于门控化上下文感知网络的词语释义生成方法,利用门控循环神经网络(GRU)对词语释义生成过程进行建模,自动为目标词生成词语释义。该模型基于编码器—解码器架构。编码器首先利用双向GRU对目标词的上下文进行编码,并采用不同的匹配策略进行目标词与上下文的交互,结合注意力机制分别从粗粒度和细粒度两个层次将上下文信息融合到目标词的向量表示中,最终获得目标词在特定语境中的编码向量。解码器则同时基于目标词的语境与语义信息为目标词生成上下文相关的词语释义。此外,通过向模型提供目标词字符级特征信息,进一步提高了生成释义的质量。在英文牛津词典数据集上进行的实验表明,该文提出的方法能够生成易于阅读和理解的词语释义,在释义建模的困惑度和生成释义的BLEU值上分别超出此前模型4.45和2.19,性能有显著提升。