基于改进Cascade R-CNN算法的船舶目标检测方法

为了解决实际场景下船舶目标检测精度低的问题,本文在Cascade R-CNN算法的基础上,提出一种船舶目标检测方法 Boat R-CNN。Boat R-CNN使用带自注意力机制的Swin-Transformer Tiny网络提取图像特征,使用Soft-NMS非极大值抑制方法提升候选框过滤精度,使用Smooth_L1损失函数加速模型收敛并减少梯度爆炸情况,使用CIOU边界框回归损失提高候选框回归质量,并针对船舶目标的形状特征优化锚框的长宽比,提高锚框的生成质量。实验结果表明,Boat R-CNN算法的精度相比原版Cascade R-CNN算法提高了21.8%,相比主流Faster R-CNN算法提高了30.3%,有效提升了实际场景下的船舶目标检测精度。

基于改进Cascade R-CNN的安全帽检测算法

针对安全帽检测中,目标形状、尺度变化大,易出现漏检、误检等问题,提出了一种基于改进级联基于区域的卷积神经网络(Cascade R-CNN)的安全帽检测算法。首先,对ResNet50进行改进形成D-ResNet50,利用可变形卷积仅增加少量参数就可增大感受野的特性,对特征提取网络的C2~C5卷积层进行重塑,提高网络对目标几何变换的适应能力和特征提取能力。其次,将D-ResNet50作为主干网络引入Cascade R-CNN,形成级联目标检测器,在每个阶段对正负样本重采样,抑制误检问题。再次,对递归特征金字塔进行改进,更高效地进行多尺度特征融合,并且基于反馈信息对特征进行二次处理,增强特征表达,提高网络的分类和定位能力。最后,使用Soft-非极大值抑制(Soft-NMS)进行后处理,进一步解决漏检问题。提出的方法在Hard hat workers数据集上的AP值相比检测基线提高了3.5%,与Sparse R-CNN、TridentNet、VFnet等先进算法相比分别提升了4.7%、5.9%、2.3%等。

基于Local Cascade Ensemble方法的胎儿健康自动分类

为更好地自动评估胎儿宫内状态,提出一种基于local cascade ensemble(LCE)方法的胎儿健康状态分类模型。选用UCI数据集,使用ADASYN方法对不平衡数据集进行填充平衡,接着结合随机森林算法对数据特征进行选择,最后使用LCE方法对胎儿状态进行自动分类。实验结果表明,所提出模型使用的方法平均准确率、精确率、召回率和F1分数分别达到了0.9554、0.9054、0.9557和0.9290,对比传统的机器学习算法能得到更好的分类效果,有效降低了误判率。

Multi-dimensional multiplexing optical secret sharing framework with cascaded liquid crystal holograms

Secret sharing is a promising technology for information encryption by splitting the secret information into different shares.However,the traditional scheme suffers from information leakage in decryption process since the amount of available information channels is limited.Herein,we propose and demonstrate an optical secret sharing framework based on the multi-dimensional multiplexing liquid crystal(LC)holograms.The LC holograms are used as spatially separated shares to carry secret images.The polarization of the incident light and the distance between different shares are served as secret keys,which can significantly improve the information security and capacity.Besides,the decryption condition is also restricted by the applied external voltage due to the variant diffraction efficiency,which further increases the information security.In implementation,an artificial neural network(ANN)model is developed to carefully design the phase distribution of each LC hologram.With the advantage of high security,high capacity and simple configuration,our optical secret sharing framework has great potentials in optical encryption and dynamic holographic display.

Pluggable multitask diffractive neural networks based on cascaded metasurfaces

Optical neural networks have significant advantages in terms of power consumption,parallelism,and high computing speed,which has intrigued extensive attention in both academic and engineering communities.It has been considered as one of the powerful tools in promoting the fields of imaging processing and object recognition.However,the existing optical system architecture cannot be reconstructed to the realization of multi-functional artificial intelligence systems simultaneously.To push the development of this issue,we propose the pluggable diffractive neural networks(P-DNN),a general paradigm resorting to the cascaded metasurfaces,which can be applied to recognize various tasks by switching internal plug-ins.As the proof-of-principle,the recognition functions of six types of handwritten digits and six types of fashions are numerical simulated and experimental demonstrated at near-infrared regimes.Encouragingly,the proposed paradigm not only improves the flexibility of the optical neural networks but paves the new route for achieving high-speed,low-power and versatile artificial intelligence systems.

Model Predictive Control for Cascaded H-Bridge PV Inverter with Capacitor Voltage Balance

We designed an improved direct-current capacitor voltage balancing control model predictive control(MPC)for single-phase cascaded H-bridge multilevel photovoltaic(PV)inverters.Compared with conventional voltage balanc-ing control methods,the method proposed could make the PV strings of each submodule operate at their maximum power point by independent capacitor voltage control.Besides,the predicted and reference value of the grid-connected current was obtained according to the maximum power output of the maximum power point tracking.A cost function was con-structed to achieve the high-precision grid-connected control of the CHB inverter.Finally,the effectiveness of the proposed control method was verified through a semi-physical simulation platform with three submodules.

基于改进Cascade RCNN的输电线路防振锤脱落检测方法

无人机巡检输电线路时,因拍摄角度和距离问题,容易出现被输电线遮挡和远距离拍摄的防振锤脱落目标,导致目标特征被遮挡且分辨率较低,且部分防振锤出现滑移现象,导致目标识别准确率降低。针对以上问题,提出一种基于改进Cascade RCNN的防振锤脱落检测网络。第一,设计了对比学习网络,将正负样本与真实样本的特征进行对比学习,利用对比损失函数训练网络,使其能更加关注到被遮挡的防振锤脱落目标,提升其特征提取能力;第二,进行了分类器增强操作,筛选出网络级联结构中回归效果较好的感兴趣区域并送入最后的分类回归队列中,提高了分类器的分类能力,进而提升检测目标的分类分数;第三,设计了并行注意力机制模块,整合网络提取的特征,增大关键特征的权重,使网络关注到图像中更关键的区域;在特征金字塔中,将双线性插值方法代替为反卷积,提升特征还原能力。经交叉验证实验结果表明,改进后的模型召回率、精确率和平均精度达到了97.5%,91.0%和92.0%,相比基线模型分别提高了6.9%,28.4%和8.0%。

改进的Cascade R-CNN算法在目标检测上的应用

针对Cascade R-CNN目标检测算法中存在检测精度较低以及目标遮挡问题,本文提出一种改进的Cas-cade R-CNN网络目标检测算法.该算法在主干网络ResNet101中引入可切换空洞卷积模块(Switchable Atrous Convolu-tion,SAC),该模块主要由两个全局上下文模块以及SAC组件构成,采用SAC组件以不同的空洞卷积率对特征进行卷积,并使用Switch函数收集特征来提高特征提取能力.同时,在ResNet101残差网络中引入坐标注意力机制(Coordi-nate Attention,CA),该机制将位置信息嵌入通道注意力中,用于更好地获取方向感知和位置感知信息,进而提高目标检测精度.此外,针对目标遮挡问题,引入Repulsion Loss损失函数.该损失函数主要由吸引项和排斥项组成,吸引项使得预测框和匹配上的目标框尽可能接近,排斥项使得预测框远离错误目标,进而减少非极大值抑制(Non-Maximum Suppression,NMS)的误检,提高目标检测中遮挡问题的检测精度.实验结果表明,在公开的科大讯飞AI挑战赛数据集上,与原算法测试性能相比,改进的Cascade R-CNN网络对该数据集检出率增长了2.39%,改进算法的识别精度有一定的提高.

基于改进Cascade R-CNN的探地雷达管线目标检测

针对人工识别探地雷达管线图像时效率低、误差大和成本高昂等问题,本文提出了一种基于改进Cascade R-CNN的管线目标智能化检测方法.首先对探地雷达管线图像数据集进行预处理,提升数据质量.然后采用ResNeXt代替ResNet作为主干网络提取目标特征信息,并添加多尺度特征融合模块FPN使高层特征向低层特征融合,增强低层特征表达能力.其次,使用高斯形式的非极大值抑制方法Soft-NMS得到更加精准的候选框,使用Smooth_L1作为损失函数,加速了模型收敛并且降低了训练中发生梯度爆炸的概率.最后,对于管线目标特殊的形状特征,设置合适的锚框长宽比和大小,提高锚框的生成质量.实验结果表明,本文方法解决了复杂特征的地下管线目标智能化检测,对地下管线目标检测的平均精度达到94.7%,比Cascade R-CNN方法提高了10.1%.

基于改进Cascade R-CNN的绝缘子故障检测方法研究

针对航拍图中存在的绝缘子故障位置在图像中占比小、背景环境复杂导致的故障检测准确率低的问题,提出一种基于改进Cascade R-CNN模型的绝缘子故障检测方法。在原有Cascade R-CNN模型的基础上,在骨干网络中引入可变形卷积学习几何变换能力,在检测器中引入平衡损失函数平衡难易样本。在模型训练阶段,使用Copy-Paste与Mosica丰富故障绝缘子样本,平衡正负样本。使用该模型对航拍绝缘子图片进行故障检测实验,改进损失函数的模型与传统Cascade R-CNN模型相比平均召回率提升0.38%,引入可变卷积后的Cascade R-CNN模型与Faster R-CNN模型,相比平均召回率,从原来的89.78%变成93.49%,结果表明该模型能够有效克服样本遮挡以及样本不平衡的干扰。

基于Cascade R-CNN的玉米幼苗检测

准确识别玉米幼苗是实现自动化精准除草、间苗、补种等苗期作业的重要前提。为此,针对自然环境下农业机器人对玉米幼苗的检测问题,结合深度残差网络强大的特征提取能力和级联网络连接多个检测器不断优化预测结果的特点,对Cascade R-CNN模型进行改进,使之适用于自然环境下玉米幼苗的检测。模型使用残差网络ResNet-50与特征金字塔网络FPN作为特征提取器提取玉米幼苗图像的特征图,利用区域建议网络生成目标候选框,通过感兴趣区域池化将不同大小的特征图转换为统一尺寸的输出;最后,分类回归模块根据特征图对目标进行分类,并使用边框回归修正候选框的位置和大小,从而完成玉米幼苗目标检测。同时,以3~5叶期玉米幼苗为研究对象,采集其田间图像并制作数据集,用所制作的数据集对Cascade R-CNN模型进行训练,选取AlexNet、VGG16、ResNet18、ResNet50与ResNet50+FPN分别作为特征提取网络进行对比试验,确定所提出的ResNet50+FPN为最优特征提取网络,平均精度均值(mAP)为91.76%,平均检测时间为6.5ms。选取双阶段目标检测模型Faster R-CNN、R-FCN、CoupleNet与以ResNet50+FPN为特征提取网络的Cascade R-CNN进行对比实验,结果表明:Cascade R-CNN模型检测效果最佳、速度最快,且能对自然环境下的玉米幼苗进行有效检测,可为玉米苗期自动化精准作业提供技术支持。

基于改进Cascade R-CNN模型的机器人抓取检测研究

为提高多物体抓取检测网络的抓取检测准确率,提出一种基于改进Cascade R-CNN模型的机械臂抓取检测算法.首先,引入ResNeXt结构能够在不加大网络设计难度的前提下提高了模型的准确率;引入带空洞卷积的空间金字塔池化模块以解决分辨率较低的问题;接着对抓取框回归分支和角度分类分支以分治方法进行优化.其次,针对多物体抓取数据集缺乏的问题,构建多目标抓取数据集(multi-object grasping dataset,MOGD),有效地扩充了多物体抓取检测数据集.最后,基于改进Cascade R-CNN模型设计抓取检测网络,实验结果表明,改进后的算法效率更高,PI-Cascade R-CNN实验准确率为93%,较Cascade R-CNN提升1.5个百分点.

基于改进Cascade R-CNN的布匹瑕疵检测算法

布匹瑕疵的自动化检测是目前纺织行业面临的一个难点问题。针对当前布匹瑕疵检测算法对尺度和长宽比变化大、小目标众多的样本检测效果并不理想的问题,提出了基于改进Cascade R-CNN网络的布匹瑕疵检测算法。首先,在特征提取网络ResNet-50中融入可变形卷积,自适应地提取更多的瑕疵形状与尺度特征;其次,在特征金字塔网络上采样前引入平衡特征金字塔,缩小特征融合前各特征层之间的语义差距,得到更具表达力的多尺度特征;然后,根据瑕疵尺度与长宽比特点重新设计更适合的初始锚框;最后,采用具有尺度不变性的GIoU Loss作为级联检测器的回归损失,以获取更加精确的瑕疵预测边界框。实验结果表明,相比基于Cascade R-CNN的算法,改进后的Cascade R-CNN算法对布匹瑕疵检测的平均精确率获得了明显提升。

Multivariable identification of membrane fouling based on compacted cascade neural network

The membrane fouling phenomenon,reflected with various fouling characterization in the membrane bioreactor(MBR)process,is so complicated to distinguish.This paper proposes a multivariable identification model(MIM)based on a compacted cascade neural network to identify membrane fouling accurately.Firstly,a multivariable model is proposed to calculate multiple indicators of membrane fouling using a cascade neural network,which could avoid the interference of the overlap inputs.Secondly,an unsupervised pretraining algorithm was developed with periodic information of membrane fouling to obtain the compact structure of MIM.Thirdly,a hierarchical learning algorithm was proposed to update the parameters of MIM for improving the identification accuracy online.Finally,the proposed model was tested in real plants to evaluate its efficiency and effectiveness.Experimental results have verified the benefits of the proposed method.

Procedure for determining the number of thermal diffusion columns in square cascade for separation of Ne stable isotopes

The thermal diffusion column represents one method of separating stable isotopes.This method is advantageous for smallscale operations because of the simplicity of the apparatus and small inventory,especially in gas-phase operations.Consequently,it has attracted attention for its applicability in tritium and noble gas separation systems.In this study,the R cascade was used to design and determine the number of columns.A square cascade was adopted for the final design because of its flexibility,and calculations were performed to separate 20Ne and 22Ne isotopes.All the R cascades that enriched the Ne isotopes by more than 99%were investigated,the number of columns was determined,and the square cascade parameters were optimized using the specified columns.Additionally,a calculation code“RSQ_CASCADE”was developed.A unit separation factor of three was considered,and the number of studied stages ranged from 10 to 20.The results showed that the column separation power,relative total flow rate,and required number of columns were linearly related to the number of stages.The separation power and relative total flow decreased and the number of columns increased as the stage number increased.Therefore,a cascade of 85 columns is recommended to separate the stable Ne isotopes.These calculations yielded a 17-stage square cascade with five columns in each stage.By changing the stage cut,feed point,and cascade feed flow rate,the best parameters for the square cascade were determined according to the cascade and column separation powers.As the column separation power had a maximum value in cascade feed 50,it was selected for separating Ne isotopes.

Anti-symmetric sampled grating quantum cascade laser for mode selection

For mode selection in a quantum cascade laser(QCL),we demonstrate an anti-symmetric sampled grating(ASG).The wavelength of the-1-th mode of this laser has been blue-shifted more than 75 nm(~10 cm^(-1))compared with that of an ordinary sampled grating laser with an emission wavelength of approximately 8.6μm,when the periodicities within both the base grating and the sample grating are kept constant.Under this condition,an improvement in the continuous tuning capability of the QCL array is ensured.The ASG structure is fabricated in holographic exposure and optical photolithography,thereby enhancing its flexibility,repeatability,and cost-effectiveness.The wavelength modulation capability of the two channels of the grating is insensitive to the variations in channel size,assuming that the overall waveguide width remains constant.The output wavelength can be tailored freely within a certain range by adjusting the width of the ridge and the material of the cladding layer.

融合注意力机制的Cascade R-CNN田间害虫检测方法

为解决测报灯采集图像中害虫依赖人工识别及统计结果可靠性低和准确性差的问题,本研究提出一种改进型Cascade R-CNN田间害虫检测算法。该算法以Cascade R-CNN为基础框架,采用ResNeSt-50作为主干网络,融合了跨通道注意力机制;使用统一目标检测头(unifying object detection heads with attentions,DyHead),并融合尺度感知、空间位置感知和任务感知。此外,采用简单复制-粘贴(simple copy-paste,SCP)方法进行了数据增强。研究共采集到20类害虫总计1 500张图像,制作了符合MS COCO格式(microsoft common objects in context 2017,MS COCO 2017)的测报灯田间害虫数据集。结果显示,本研究提出的方法的F1分数(F1-score)达到了86.2%。当交并比(intersection over union,IoU)为0.5时,其F1-分数与经典Cascade R-CNN、Faster R-CNN和YOLOv4相比,分别提升了2.8、5.8和8.2个百分点。表明该方法满足测报灯害虫检测任务对判别能力和实时性的要求,实现了害虫的高精度自动识别与计数,可直接应用于田间害虫检测。

Quantum cascade lasers grown by MOCVD

Sharing the advantages of high optical power,high efficiency and design flexibility in a compact size,quantum cascade lasers(QCLs)are excellent mid-to-far infrared laser sources for gas sensing,infrared spectroscopic,medical diagnosis,and defense applications.Metalorganic chemical vapor deposition(MOCVD)is an important technology for growing high quality semiconductor materials,and has achieved great success in the semiconductor industry due to its advantages of high efficiency,short maintenance cycles,and high stability and repeatability.The utilization of MOCVD for the growth of QCL materials holds a significant meaning for promoting the large batch production and industrial application of QCL devices.This review summarizes the recent progress of QCLs grown by MOCVD.Material quality and the structure design together determine the device performance.Research progress on the performance improvement of MOCVD-grown QCLs based on the optimization of material quality and active region structure are mainly reviewed.

基于Cascade RCNN的热轧带钢表面缺陷检测

在工业生产过程中,带钢表面产生的缺陷影响其质量和使用性能,需要对表面缺陷进行检测,为此提出了基于深度学习的缺陷检测模型Cascade RCNN。首先,对骨干网络进行改进,将标准卷积替换为可切换空洞卷积,在不增加参数量的情况下,增大输出单元的感受野。其次,改变特征金字塔FPN,结构不变的情况下添加了自上而下的连接方式,同时使用特征上采样算子CARAFE替换最邻近上采样,提高了上采样精度和定位精度。最后,将损失函数换为Focal Loss,解决目标检测过程中正负样本不平衡问题。结果显示:通过以上方法的改进,检测精度有大幅提升,平均均值精度提高了7.61%,达到77.82%,各类缺陷的检测精度都得到了提高;与其他检测模型对比,模型的检测能力得到了提高,采用的改进方法有一定的应用价值。

A Detection Method of Bolts on Axlebox Cover Based on Cascade Deep Convolutional Neural Network

This paper proposes a cascade deep convolutional neural network to address the loosening detection problem of bolts on axlebox covers.Firstly,an SSD network based on ResNet50 and CBAM module by improving bolt image features is proposed for locating bolts on axlebox covers.And then,theA2-PFN is proposed according to the slender features of the marker lines for extracting more accurate marker lines regions of the bolts.Finally,a rectangular approximationmethod is proposed to regularize themarker line regions asaway tocalculate the angle of themarker line and plot all the angle values into an angle table,according to which the criteria of the angle table can determine whether the bolt with the marker line is in danger of loosening.Meanwhile,our improved algorithm is compared with the pre-improved algorithmin the object localization stage.The results show that our proposed method has a significant improvement in both detection accuracy and detection speed,where ourmAP(IoU=0.75)reaches 0.77 and fps reaches 16.6.And in the saliency detection stage,after qualitative comparison and quantitative comparison,our method significantly outperforms other state-of-the-art methods,where our MAE reaches 0.092,F-measure reaches 0.948 and AUC reaches 0.943.Ultimately,according to the angle table,out of 676 bolt samples,a total of 60 bolts are loose,69 bolts are at risk of loosening,and 547 bolts are tightened.