Point cloud upsampling generative adversarial network based on residual multi-scale off-set attention

Background Owing to the limitations of the working principle of three-dimensional(3D) scanning equipment, the point clouds obtained by 3D scanning are usually sparse and unevenly distributed. Method In this paper, we propose a new generative adversarial network(GAN) that extends PU-GAN for upsampling of point clouds. Its core architecture aims to replace the traditional self-attention(SA) module with an implicit Laplacian offset attention(OA) module and to aggregate the adjacency features using a multiscale offset attention(MSOA)module, which adaptively adjusts the receptive field to learn various structural features. Finally, residual links are added to create our residual multiscale offset attention(RMSOA) module, which utilizes multiscale structural relationships to generate finer details. Result The results of several experiments show that our method outperforms existing methods and is highly robust.

Scheme Based on Multi-Level Patch Attention and Lesion Localization for Diabetic Retinopathy Grading

Early screening of diabetes retinopathy(DR)plays an important role in preventing irreversible blindness.Existing research has failed to fully explore effective DR lesion information in fundus maps.Besides,traditional attention schemes have not considered the impact of lesion type differences on grading,resulting in unreasonable extraction of important lesion features.Therefore,this paper proposes a DR diagnosis scheme that integrates a multi-level patch attention generator(MPAG)and a lesion localization module(LLM).Firstly,MPAGis used to predict patches of different sizes and generate a weighted attention map based on the prediction score and the types of lesions contained in the patches,fully considering the impact of lesion type differences on grading,solving the problem that the attention maps of lesions cannot be further refined and then adapted to the final DR diagnosis task.Secondly,the LLM generates a global attention map based on localization.Finally,the weighted attention map and global attention map are weighted with the fundus map to fully explore effective DR lesion information and increase the attention of the classification network to lesion details.This paper demonstrates the effectiveness of the proposed method through extensive experiments on the public DDR dataset,obtaining an accuracy of 0.8064.

Deep neural network based on multi-level wavelet and attention for structured illumination microscopy

Structured illumination microscopy(SIM)is a popular and powerful super-resolution(SR)technique in biomedical research.However,the conventional reconstruction algorithm for SIM heavily relies on the accurate prior knowledge of illumination patterns and signal-to-noise ratio(SNR)of raw images.To obtain high-quality SR images,several raw images need to be captured under high fluorescence level,which further restricts SIM’s temporal resolution and its applications.Deep learning(DL)is a data-driven technology that has been used to expand the limits of optical microscopy.In this study,we propose a deep neural network based on multi-level wavelet and attention mechanism(MWAM)for SIM.Our results show that the MWAM network can extract high-frequency information contained in SIM raw images and accurately integrate it into the output image,resulting in superior SR images compared to those generated using wide-field images as input data.We also demonstrate that the number of SIM raw images can be reduced to three,with one image in each illumination orientation,to achieve the optimal tradeoff between temporal and spatial resolution.Furthermore,our MWAM network exhibits superior reconstruction ability on low-SNR images compared to conventional SIM algorithms.We have also analyzed the adaptability of this network on other biological samples and successfully applied the pretrained model to other SIM systems.

Mu-Net:Multi-Path Upsampling Convolution Network for Medical Image Segmentation

Medical image segmentation plays an important role in clinical diagnosis,quantitative analysis,and treatment process.Since 2015,U-Net-based approaches have been widely used formedical image segmentation.The purpose of the U-Net expansive path is to map low-resolution encoder feature maps to full input resolution feature maps.However,the consecutive deconvolution and convolutional operations in the expansive path lead to the loss of some high-level information.More high-level information can make the segmentationmore accurate.In this paper,we propose MU-Net,a novel,multi-path upsampling convolution network to retain more high-level information.The MU-Net mainly consists of three parts:contracting path,skip connection,and multi-expansive paths.The proposed MU-Net architecture is evaluated based on three different medical imaging datasets.Our experiments show that MU-Net improves the segmentation performance of U-Net-based methods on different datasets.At the same time,the computational efficiency is significantly improved by reducing the number of parameters by more than half.

FPGA based hardware platform for trapped-ion-based multi-level quantum systems

We report a design and implementation of a field-programmable-gate-arrays(FPGA)based hardware platform,which is used to realize control and signal readout of trapped-ion-based multi-level quantum systems.This platform integrates a four-channel 2.8 Gsps@14 bits arbitrary waveform generator,a 16-channel 1 Gsps@14 bits direct-digital-synthesisbased radio-frequency generator,a 16-channel 8 ns resolution pulse generator,a 10-channel 16 bits digital-to-analogconverter module,and a 2-channel proportion integration differentiation controller.The hardware platform can be applied in the trapped-ion-based multi-level quantum systems,enabling quantum control of multi-level quantum system and highdimensional quantum simulation.The platform is scalable and more channels for control and signal readout can be implemented by utilizing more parallel duplications of the hardware.The hardware platform also has a bright future to be applied in scaled trapped-ion-based quantum systems.

Deep reinforcement learning based multi-level dynamic reconfiguration for urban distribution network:a cloud-edge collaboration architecture

With the construction of the power Internet of Things(IoT),communication between smart devices in urban distribution networks has been gradually moving towards high speed,high compatibility,and low latency,which provides reliable support for reconfiguration optimization in urban distribution networks.Thus,this study proposed a deep reinforcement learning based multi-level dynamic reconfiguration method for urban distribution networks in a cloud-edge collaboration architecture to obtain a real-time optimal multi-level dynamic reconfiguration solution.First,the multi-level dynamic reconfiguration method was discussed,which included feeder-,transformer-,and substation-levels.Subsequently,the multi-agent system was combined with the cloud-edge collaboration architecture to build a deep reinforcement learning model for multi-level dynamic reconfiguration in an urban distribution network.The cloud-edge collaboration architecture can effectively support the multi-agent system to conduct“centralized training and decentralized execution”operation modes and improve the learning efficiency of the model.Thereafter,for a multi-agent system,this study adopted a combination of offline and online learning to endow the model with the ability to realize automatic optimization and updation of the strategy.In the offline learning phase,a Q-learning-based multi-agent conservative Q-learning(MACQL)algorithm was proposed to stabilize the learning results and reduce the risk of the next online learning phase.In the online learning phase,a multi-agent deep deterministic policy gradient(MADDPG)algorithm based on policy gradients was proposed to explore the action space and update the experience pool.Finally,the effectiveness of the proposed method was verified through a simulation analysis of a real-world 445-node system.

Neighborhood Effects and Political Trust: A Multi-level Analysis of Chinese Rural-to-Urban Migrants’ Trust in County Government

Massive rural-to-urban migration in China is consequential for political trust: rural-to-urban migrants have been found to hold lower levels of trust in local government than their rural peers who choose to stay in the countryside (mean 4.92 and 6.34 out of 10, respectively, p < 0.001). This article explores why migrants have a certain level of political trust in their county-level government. Using data of rural-to-urban migrants from the China Family Panel Survey, this study performs a hierarchical linear modeling (HLM) to unpack the multi-level explanatory factors of rural-to-urban migrants’ political trust. Findings show that the individual-level socio-economic characteristics and perceptions of government performance (Level-1), the neighborhood-level characteristics-the physical and social status and environment of neighborhoods (Level-2), and the objective macroeconomic performance of county-level government (Level-3), work together to explain migrants’ trust levels. These results suggest that considering the effects of neighborhood-level factors on rural-to-urban migrants’ political trust merits policy and public management attention in rapidly urbanizing countries.

基于改进YOLOv5s的小目标检测算法

针对当前主流目标检测算法对图像中远距离小目标产生的漏检、误检等问题,提出一种改进YOLOv5s的小目标检测算法。在模型训练过程中,通过引入Focal-EIOU定位损失函数,加强边界框的定位精度;在骨干网络中,通过添加小目标检测层,提高小目标的检测精度;在Neck结构中,通过优化上采样算子和添加注意力机制,加强小目标的特征信息。实验结果表明,改进后的算法在VisDrone数据集上与YOLOv5s算法相比,mAP@small提高了3.2%,且检测速度满足实时性的要求,能够很好地应用于小目标检测任务中。

基于注意力与密集重参数化的目标检测算法

针对目标检测任务中背景复杂、目标尺寸差异大等因素导致目标检测结果较差的问题,本文提出基于注意力和密集重参数化的目标检测算法。首先,基于CSP-DarkNet提出高效的特征提取网络,主要包括密集重参数化模块和CASA模块2个设计。前者利用密集连接保留浅层特征,又通过重参数化结构降低网络复杂度;后者CASA模块用于获取需要的目标信息。其次,特征融合在特征金字塔(FPN)和路径聚合网络(PAN)的基础上,引入内容感知特征重组(CARAFE)进行上采样,有效解决了邻近插值法等未能捕捉丰富语义信息的问题;提出更高效的C3-G模块,获取丰富的梯度信息,增强模型表达能力和感知能力;同时,引入深度可分离卷积提升运算效率。最后,检测输出采用在更大范围上跨领域正负样本匹配策略扩充正样本数量,提升检测效果。该算法在MS COCO和PASCAL VOC数据集上的mAP@0.5分别达到了57.5%和83.0%,充分说明了本文算法的先进性。

DFNet:高效的无解码语义分割方法

针对编解码语义分割网络计算量大、解码结构复杂的问题,提出一种高效无解码的二值语义分割模型DFNet。该模型首先去除主流分割网络中复杂的解码结构和跳跃连接,采用卷积重塑上采样方法重塑特征编码直接得到分割结果,简化网络模型结构;其次在编码器中融合轻量双重注意力机制EC&SA,提高特征编码的通道及空间信息交互,增强网络的编码能力;最后使用PolyCE损失替代常规分割损失,解决正负样本不均衡问题,提高模型的分割精度。在Deep‑Globe道路分割和CrackForest缺陷检测等二值分割数据集上的实验结果表明,本文模型的分割精度F1均值和IoU均值分别达到84.69%和73.95%,且分割速度高达94 FPS,远超主流语义分割模型,极大地提高了分割任务效率。

用于数据采集器测试的时间脉冲插值分析方法

地震数据采集器时间服务精度是评价设备性能的重要指标,一般使用标准时间源输出的时间脉冲信号采样数据进行分析计算,但受仪器采样率限制,时间偏差测试分辨率不高。设计了基于一阶差分和直方图分析的脉冲信号上升沿识别、高低电位分析和时间偏差分析算法,研究使用非线性插值分析方法对地震数据采集器采集的标准时间脉冲信号进行波形升采样恢复,并基于所提算法进行实验分析和讨论,在提高地震数据采集器时间偏差测试分辨率的同时,规范了测试数据处理方法和流程。

基于改进DeeplabV3+的遥感图像道路分割模型

针对遥感图像道路分割边界模糊和遮挡难以区分的问题,提出了基于改进DeeplabV3+的遥感图像道路分割模型。该模型在主干网络中引入MobileNetV3和高效通道注意力机制(ECA),减少了参数量并关注连续的道路特征信息。在解码过程中采用多级上采样,增强了编码器和解码器之间的紧密连接,全面保留了细节信息。同时,在ASPP模块中采用深度可分离膨胀卷积DS-ASPP,显著减少了参数量。实验结果表明,该模型在Massachusetts Roads数据集上的交并比达到了83.71%,准确率达到了93.71%,分割精度最优,模型参数量为55.57×10^(6),能够有效地避免边界模糊和遮挡导致的错漏检问题,在遥感道路分割中提高了精度和速度。

基于点代理增强和逐层上采样的猪体点云补全方法

采用逆向工程技术进行猪体的三维重建并测算,是低成本无接触式猪体型体况评估的一大解决方案,在比较单视角和多视角采集方法的优缺点后,本文提出基于深度学习的点云补全方法,将猪体局部点云恢复成一个完整的点云以实现猪体三维重建。该猪体点云补全方法基于点代理增强和逐层上采样,首先通过特征提取结合位置嵌入生成点代理,使用点代理增强Transformer进一步提高点代理的特征表示能力,再基于点代理通过逐层上采样由粗到细逐步恢复最终的高分辨率、细粒度和分布均匀的完整点云。本文对实际生产环境中采集的猪体点云进行补全,所提方法与目前主流的点云补全方法进行对比试验,在多个指标的评定上,本文提出的方法都取得了较好性能,尤其是在猪体点云缺失严重补全难度较大的情况下效果更为突出。试验证明该方法对猪体主干部位的补全具备应用价值,能够用于实现基于局部点云的猪体三维点云重建。

多尺度特征金字塔融合的街景图像语义分割

针对街景图像语义分割任务中的目标尺寸差异大、多尺度特征难以高效提取的问题,本文提出了一种语义分割网络(LDPANet).首先,将空洞卷积与引入残差学习单元的深度可分离卷积结合,来优化编码器结构,在降低了计算复杂度的同时缓解梯度消失的问题.然后利用层传递的迭代空洞空间金字塔,将自顶向下的特征信息依次融合,提高了上下文信息的有效交互能力;在多尺度特征融合之后引入属性注意力模块,使网络抑制冗余信息,强化重要特征.再者,以通道扩展上采样代替双线插值上采样作为解码器,进一步提升了特征图的分辨率.最后,LDPANet方法在Cityscapes和CamVid数据集上的精度分别达到了91.8%和87.52%,与近几年网络模型相比,本文网络模型可以精确地提取像素的位置信息以及空间维度信息,提高了语义分割的准确率.

改进YOLOv5s的铁路异物入侵检测算法

行人、车辆等异物侵入铁路界线,严重威胁行人安全和铁路交通安全.针对传统铁路异物入侵检测方法精度低、时效性差等问题,提出改进YOLOv5s算法的铁路异物入侵检测模型SD-YOLO.本文提出SSA混合注意力机制,加强模型的局部表征能力,提高小目标识别效果;提出DW-Decoupled Head解耦检测头,加快网络收敛速度;引入边界框回归损失函数SIoU,提高了模型的检测精度;使用转置卷积作为采样方法,采样更适合铁路侵限障碍物特征的尺寸和比例.在数据集RS和Pascal VOC 2012进行实验验证,与基线YOLOv5s算法相比,平均精度mAP@0.5分别提高了2.7%、1.8%,mAP@.5:.95分别提高了2.9%、2.1%,检测速度分别达到79 FPS和78 FPS,表明该算法在检测精度和速度上均取得良好的性能,有效改善了漏检、误检问题,提高了小目标识别能力.

基于BiLevelNet的实时语义分割算法

针对语义分割网络参数量过大导致其难以部署在内存受限的边缘设备等问题,本文提出一种基于BiLevelNet的轻量级实时语义分割算法。首先,利用空洞卷积扩大感受野,并结合特征复用策略增强网络的区域感知能力。接着,嵌入两阶段的PBRA注意力机制,建立远距离相关物体之间的依赖关系以增强网络的全局感知能力。最后,引入结合浅层特征的FADE算子以改善图像上采样效果。实验结果表明,在输入图像分辨率为512×1024的情况下,本文网络在Cityscapes数据集上以121 f/s的速率获得了75.1%的平均交并比,模型大小仅为0.7 M。同时在输入图像分辨率为360×480的情况下,在Camvid数据集上取得68.2%的平均交并比。同当前其他实时语义分割方法相比,该网络性能取得速度与精度的均衡,符合自动驾驶应用场景对实时性的要求。

基于DFECANet的遥感图像飞机目标检测方法

针对现有的遥感图像目标检测方法中对小尺寸飞机目标的检测精度不高、特征信息传递不准确、信息交互不充分等问题,提出了一种基于可辨别特征提取和上下文感知的遥感图像飞机目标检测方法。设计了以可辨别特征提取模块为主体的主干网络,用以加强对多尺度飞机目标的特征提取;引入自适应特征增强模块,选择性关注小目标、优化特征信息的传递与信息交互;并设计了特征融合上采样模块对特征图进行上采样操作,用以提升高层语义信息的准确性。在DOTAv1数据集上的检测精度达到了95.2%,相较于YOLOv5s、SCRDet、ASSD等主流算法,飞机目标的检测精度提高了3.7%~18%。此外,该方法的检测速度以及模型参数量分别为147 fps和13.4 M,相较于当前主流算法具备较强的竞争力,满足在遥感背景下对飞机目标的实时检测需求。

一种面向行人跌倒检测的改进YOLOv5算法

针对行人跌倒检测的应用场景复杂,原始YOLOv5的检测精度不高,特征融合不充分的问题,本文首先提出了一种新的特征金字塔(FPN)结构-自适应特征增强融合金字塔网络(AFEF-FPN)和特征增强融合模块(FFEM)来增加特征的表示以及融合深度.其次使用基于全局语义信息的上采样算子CARAFE代替Nearest Upsample来建模全局上采样信息,使用Alpha IoU Loss代替CIoU Loss来提高High IoU目标的损失和回归精度.最终本文网络YOLOv5(AFEF-FPN)在训练集上取得了98.62%mAP,在测试集上取得了96.21%mAP,相比于原始YOLOv5网络在训练集和测试集上分别提升了1.64%和2.86%.实验表明,本文网络在复杂场景下的目标检测效果优于原始YOLOv5及其他网络.

DGOA:基于动态图和偏移注意力的点云上采样

由三维扫描设备直接得到的点云经常是稀疏、不均匀、有噪声的,因而点云上采样在点云重建、渲染等领域扮演了越来越关键的角色。为此提出了一种新的基于动态图和偏移注意力的点云上采样网络DGOA,主要包含局部特征提取(LFE)、全局特征提取(GFE)和坐标重建(CR)3个模块。LFE采用多层结构提取邻域信息,每层基于特征相似性构建动态图,可以在特征空间自适应的将点云分组,增大感受野,获得长距离的语义信息,更好的建模点云的局部几何形状。GFE采用基于拉普拉斯算子的偏移注意力使每个点都能获得点云的全局信息,使生成点云的细节与原始点云一致,减少噪声的影响。CR借鉴FoldingNet操作,避免生成点的聚集。此外,整个网络与输入点云中点的顺序无关,具有置换不变性。在多个数据集的定量与定性实验结果表明,该方法优于其他方法,并且具有良好的泛化性和稳定性。

基于改进YOLOv5的轨道交通障碍物检测算法

针对复杂的轨道交通背景下障碍物检测精度低和检测速度慢的问题,提出了一种改进YOLOv5的目标检测网络模型。首先,采用基于注意力机制的轻量级Transformer主干EMO来替换YOLOv5原有backbone中的部分模块,保证轻量化的同时,还能够提高模型的准确性和稳定性;其次,使用Focal-EIoU来替换YOLOv5中的CIoU损失函数,以解决CIoU引起的训练效率低、收敛速度慢等问题;最后使用轻量化上采样算子CARAFE来替换YOLOv5算法中原有的上采样层,在没有引入过多参数和计算量的情况下具有更大的感受野,提高了检测精度和检测速度。实验结果表明,该方法相较于原始的YOLOv5网络模型平均精确度提升了11.1%,准确率提升了13%,召回率提升了11.4%,检测速度达到了60.7 fps。所提出的方法在目标检测任务中表现出了较好的性能,有效增强了轨道交通背景下目标检测模型的检测性能。