"计算机网络基础"课程"金课"建设初探

为了响应教育部关于建设"金课"的倡议,基于我院计算机科学与技术专业的"计算机网络基础"课程特点,结合该专业的毕业要求,修订了课程目标;根据课程目标,重新整合了教学内容,确定了线上线下混合式教学模式;构建课程目标达成度评价体系,并将其用于后续教学过程.实践结果表明,该教学改革符合混合式"金课"的建设要求,提升了教学效果.     

改进R-FCN模型的小尺度行人检测

目的 为了有效解决传统行人检测算法在分辨率低、行人尺寸较小等情境下检测精度低的问题,将基于区域全卷积网络（region-based fully convolutional networks,R-FCN）的目标检测算法引入到行人检测中,提出一种改进R-FCN模型的小尺度行人检测算法。方法 为了使特征提取更加准确,在ResNet-101的conv5阶段中嵌入可变形卷积层,扩大特征图的感受野;为提高小尺寸行人检测精度,在ResNet-101中增加另一条检测路径,对不同尺寸大小的特征图进行感兴趣区域池化;为解决小尺寸行人检测中的误检问题,利用自举策略的非极大值抑制算法代替传统的非极大值抑制算法。结果 在基准数据集Caltech上进行评估,实验表明,改进的R-FCN算法与具有代表性的单阶段检测器（single shot multiBox detector,SSD）算法和两阶段检测器中的Faster R-CNN（region convolutional neural network）算法相比,检测精度分别提高了3.29%和2.78%;在相同ResNet-101基础网络下,检测精度比原始R-FCN算法提高了12.10%。结论 本文提出的改进R-FCN模型,使小尺寸行人检测精度更加准确。相比原始模型,改进的R-FCN模型对行人检测的精确率和召回率有更好的平衡能力,在保证精确率的同时,具有更大的召回率。     

Seg-CapNet：心脏MRI图像分割神经网络模型

目的 针对现有神经网络模型需要对左心室心肌内膜和外膜单独建模的问题,本文提出了一种基于胶囊结构的心脏磁共振图像（magnetic resonance imaging,MRI）分割模型Seg-CapNet,旨在同时提取心肌内膜和外膜,并保证两者的空间位置关系。方法 首先利用胶囊网络将待分割目标转换成包含目标相对位置、颜色以及大小等信息的向量,然后使用全连接将这些向量的空间关系进行重组,最后采用反卷积对特征图进行上采样,将分割图还原为输入图像尺寸。在上采样过程中将每层特征图与卷积层的特征图进行连接,有助于图像细节还原以及模型的反向传播,加快训练过程。Seg-CapNet的输出向量不仅有图像的灰度、纹理等底层图像特征,还包含目标的位置、大小等语义特征,有效提升了目标图像的分割精度。为了进一步提高分割质量,还提出了一种新的损失函数用于约束分割结果以保持多目标区域间的相对位置关系。结果 在ACDC（automated cardiac diagnosis challenge）2017、MICCAI（medical image computing and computer-assisted intervention）2013和MICCAI2009等3个心脏MRI分割竞赛的公开数据集上对Seg-CapNet模型进行训练和验证,并与神经网络分割模型U-net和SegNet进行对比。实验结果表明,相对于U-Net和SegNet,Seg-CapNet同时分割目标重叠区域的平均Dice系数提升了3.5%,平均豪斯多夫距离（Hausdorff distance,HD）降低了18%。并且Seg-CapNet的参数量仅为U-Net的54%、SegNet的40%,在提升分割精度的同时,降低了训练时间和复杂度。结论 本文提出的Seg-CapNet模型在保证同时分割重叠区域目标的同时,降低了参数量,提升了训练速度,并保持了较好的左心室心肌内膜和外膜分割精度。     

A low‐profile antenna with circularly polarized reconfigurable and omnidirectional radiation patterns

A low profile circularly polarized (CP) antenna with reconfigurable polarization is designed and presented, which can radiate omnidirectional patterns that can be switched between left‐hand circularly polarized (LHCP) and right‐hand circularly polarized (RHCP). A pair of arc‐shaped complementary dipoles is acted as reconfigurable elements by bridging four pin diodes at the dipole arced arms. A meander phase shift line is employed to connected the arc dipole arms and plate cavity to adjust the phase relationship of two sources. The proposed antenna exhibits the omnidirectional radiation pattern by combining six identical elements placed in a circular array configuration. 24 p‐i‐n diodes are exploited to six elements, by manipulating the dc bias voltage across the diodes, the polarization state of the antenna can be switched. The patterns of the antenna are similar to that of a dipole, but its size is only about Φ0.87 × 0.029λ₀ at 5.8 GHz. The overlapped bandwidth of measured 3‐dB axial ratio (AR) and 10‐dB return loss is 5.724‐5.87 and 5.738‐5.91 GHz for two polarization states, which are right on the target of ISM band. It can be well adapted to medical diagnosis systems.     

An automatic 2D to 3D video conversion approach based on RGB-D images

3D movies/videos have become increasingly popular in the market; however, they are usually produced by professionals. This paper presents a new technique for the automatic conversion of 2D to 3D video based on RGB-D sensors, which can be easily conducted by ordinary users. To generate a 3D image, one approach is to combine the original 2D color image and its corresponding depth map together to perform depth image-based rendering (DIBR). An RGB-D sensor is one of the inexpensive ways to capture an image and its corresponding depth map. The quality of the depth map and the DIBR algorithm are crucial to this process. Our approach is twofold. First, the depth maps captured directly by RGB-D sensors are generally of poor quality because there are many regions missing depth information, especially near the edges of objects. This paper proposes a new RGB-D sensor based depth map inpainting method that divides the regions with missing depths into interior holes and border holes. Different schemes are used to inpaint the different types of holes. Second, an improved hole filling approach for DIBR is proposed to synthesize the 3D images by using the corresponding color images and the inpainted depth maps. Extensive experiments were conducted on different evaluation datasets. The results show the effectiveness of our method.

     

Robust person re-identification via graph convolution networks

Person re-identification (re-id) aims to identity the same person over multiple cameras; it has been successfully applied to various computer vision applications as a fundamental method. Owing to the development of deep learning, person re-id methods, which typically use triplet networks based on triplet loss, have demonstrated great success. However, the appearances of people are similar and hence difficult to distinguish in many cases. Therefore, we present a novel graph convolution network and enhances traditional triplet loss functions. Our method defines reference, positive, and negative features for triplet loss as three vertices of a graph, respectively, and adjusts their mutual distance through learning. The method adopts graph convolutions efficiently, thereby affording low computational costs. Experimental results demonstrate that our method is superior to the baseline on the Market-1501 dataset. The proposed GCN-based triplet loss considerably contributes to improve re-identification methods quantitatively and qualitatively.

     

面向隐私保护联邦学习的医学影像目标检测算法

联邦学习技术是一种新型多机构协同训练模型范式,广泛应用于多领域,其中模型参数隐私保护是一个关键问题.针对CT影像综合性病灶检测任务,提出隐私保护的联邦学习算法.首先部署松散耦合的客户端-服务器架构;其次在各客户端使用改进的RetinaNet检测器,引入上下文卷积和后向注意力机制;最后完成联邦训练.各客户端使用局部更新策略,采用自适应训练周期,局部目标函数中加入了限制项;服务器使用自适应梯度裁剪策略和高斯噪声差分隐私算法更新全局模型参数.在DeepLesion数据集上的消融分析说明了算法各部分的重要性.实验结果表明,改进的RetinaNet检测器有效地提升了多尺度病灶的检测精度.与集中数据训练模型范式相比,联邦学习所得模型性能略低(mAP分别为75.33％和72.80％),但训练用时缩短近38％,有效地实现了隐私保护、通信效率和模型性能的良好权衡.     

自适应多模态特征融合胶质瘤分级网络

目的 胶质瘤的准确分级是辅助制定个性化治疗方案的主要手段,但现有研究大多数集中在基于肿瘤区域的分级预测上,需要事先勾画感兴趣区域,无法满足临床智能辅助诊断的实时性需求。因此,本文提出一种自适应多模态特征融合网络（adaptive multi-modal fusion net,AMMFNet）,在不需要勾画肿瘤区域的情况下,实现原始采集图像到胶质瘤级别的端到端准确预测。方法 AMMFNet方法采用4个同构异义网络分支提取不同模态的多尺度图像特征;利用自适应多模态特征融合模块和降维模块进行特征融合;结合交叉熵分类损失和特征嵌入损失提高胶质瘤的分类精度。为了验证模型性能,本文采用MICCAI （Medical Image Computing and Computer Assisted Intervention Society）2018公开数据集进行训练和测试,与前沿深度学习模型和最新的胶质瘤分类模型进行对比,并采用精度以及受试者曲线下面积（area under curve,AUC）等指标进行定量分析。结果 在无需勾画肿瘤区域的情况下,本文模型预测胶质瘤分级的AUC为0.965;在使用肿瘤区域时,其AUC高达0.997,精度为0.982,比目前最好的胶质瘤分类模型——多任务卷积神经网络同比提高1.2%。结论 本文提出的自适应多模态特征融合网络,通过结合多模态、多语义级别特征,可以在未勾画肿瘤区域的前提下,准确地实现胶质瘤分级预测。     

RLINK: Deep reinforcement learning for user identity linkage

World Wide Web - User identity linkage is a task of recognizing the identities of the same user across different social networks (SN). Previous works tackle this problem via estimating the pairwise...     

一种辅助新型冠状病毒肺炎检测的肺实质分割算法

新型冠状病毒肺炎给人类健康及社会经济造成了巨大的负面影响,而X光胸片中的肺实质提取成为新型冠状病毒肺炎诊断过程中的关键环节。在U-Net的基础上,提出一种结合编解码模式的肺实质分割算法。应用特征融合思想,构建A形特征融合模块,充分学习深层特征的语义信息。引入注意力机制,在深层卷积神经网络中加入密集空洞卷积模块和残差多核池化模块,扩大卷积感受野并提取上下文特征信息。通过改进可变形卷积和分割损失函数,提升网络模型的泛化能力和鲁棒性。实验结果表明,该算法的分割准确度、Dice系数、敏感度、Jaccard指数分别为98.16%、98.32%、98.13%、98.54%,能够实现X光胸片中肺实质部位的有效分割。