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.

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.

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.

基于改进的Cascade RCNN铸管字符检测算法

由于工业现场采集的铸管字符图像存在背景模糊、字符区域占比小、刻字位置不固定、油漆遮挡等问题,导致现有模型的检测精度难以满足工业现场的需求。针对上述问题,提出改进的Cascade RCNN铸管字符检测算法。首先对特征金字塔进行改进,提出融合小目标增强的特征金字塔(STE-FPN),利用多尺度特征融合的特征增强能力丰富铸管小目标字符的特征信息。其次引入自矫正/池化的ResNeSt(SCP-ResNeSt)作为特征提取网络,利用自矫正卷积和池化操作以提升背景复杂的铸管字符特征提取效率。最后对级联结构进行改进,引进Mask分支结构,可以自适应地检测字符区域并去除干扰区域,优化了检测结果。将改进后的算法在铸管数据集上进行测试,其平均检测精度mAP为99.1%,比原Cascade RCNN算法提高了2.3%,得到的精度表明改进后的性能优于原算法。

Using Genetic Algorithms to Improve the Search of the Weight Space in Cascade－Correlation Neural Network

In this paper, we use the global search characteristics of genetic algorithms to help search the weight space of the neurons in the cascade-correlation architecture. The cascade-correlation learning architecture is a technique of training and building neural networks that starts with a simple network of neurons and adds additional neurons as they are needed to suit a particular problem. In our approach, instead ofmodifying the genetic algorithm to account for convergence problems, we search the weight-space using the genetic algorithm and then apply the gradient technique of Quickprop to optimize the weights. This hybrid algorithm which is a combination of genetic algorithms and cascade-correlation is applied to the two spirals problem. We also use our algorithm in the prediction of the cyclic oxidation resistance of Ni- and Co-base superalloys.

A CASCADED MODEL OF NEURAL NETWORK FOR PATTERN RECOGNITION

A cascaded model of neural network and its learning algorithm suitable for opticalimplementation are proposed.Computer simulations have shown that this model may successfullybe applied to an error-tolerance pattern recognitions of multiple 3-D targets with arbitrary spatialorientations.

3D laser scanning strategy based on cascaded deep neural network

A 3D laser scanning strategy based on cascaded deep neural network is proposed for the scanning system converted from 2D Lidar with a pitching motion device. The strategy is aimed at moving target detection and monitoring. Combining the device characteristics, the strategy first proposes a cascaded deep neural network, which inputs 2D point cloud, color image and pitching angle. The outputs are target distance and speed classification. And the cross-entropy loss function of network is modified by using focal loss and uniform distribution to improve the recognition accuracy. Then a pitching range and speed model are proposed to determine pitching motion parameters. Finally, the adaptive scanning is realized by integral separate speed PID. The experimental results show that the accuracies of the improved network target detection box, distance and speed classification are 90.17%, 96.87% and 96.97%, respectively. The average speed error of the improved PID is 0.4239°/s, and the average strategy execution time is 0.1521 s.The range and speed model can effectively reduce the collection of useless information and the deformation of the target point cloud. Conclusively, the experimental of overall scanning strategy show that it can improve target point cloud integrity and density while ensuring the capture of target.

All-optical computing based on convolutional neural networks

The rapid development of information technology has fueled an ever-increasing demand for ultrafast and ultralow-en-ergy-consumption computing.Existing computing instruments are pre-dominantly electronic processors,which use elec-trons as information carriers and possess von Neumann architecture featured by physical separation of storage and pro-cessing.The scaling of computing speed is limited not only by data transfer between memory and processing units,but also by RC delay associated with integrated circuits.Moreover,excessive heating due to Ohmic losses is becoming a severe bottleneck for both speed and power consumption scaling.Using photons as information carriers is a promising alternative.Owing to the weak third-order optical nonlinearity of conventional materials,building integrated photonic com-puting chips under traditional von Neumann architecture has been a challenge.Here,we report a new all-optical comput-ing framework to realize ultrafast and ultralow-energy-consumption all-optical computing based on convolutional neural networks.The device is constructed from cascaded silicon Y-shaped waveguides with side-coupled silicon waveguide segments which we termed“weight modulators”to enable complete phase and amplitude control in each waveguide branch.The generic device concept can be used for equation solving,multifunctional logic operations as well as many other mathematical operations.Multiple computing functions including transcendental equation solvers,multifarious logic gate operators,and half-adders were experimentally demonstrated to validate the all-optical computing performances.The time-of-flight of light through the network structure corresponds to an ultrafast computing time of the order of several picoseconds with an ultralow energy consumption of dozens of femtojoules per bit.Our approach can be further expan-ded to fulfill other complex computing tasks based on non-von Neumann architectures and thus paves a new way for on-chip all-optical computing.

Inner Cascaded U^(2)-Net:An Improvement to Plain Cascaded U-Net

Deep neural networks are now widely used in the medical image segmentation field for their performance superiority and no need of manual feature extraction.U-Net has been the baseline model since the very beginning due to a symmetricalU-structure for better feature extraction and fusing and suitable for small datasets.To enhance the segmentation performance of U-Net,cascaded U-Net proposes to put two U-Nets successively to segment targets from coarse to fine.However,the plain cascaded U-Net faces the problem of too less between connections so the contextual information learned by the former U-Net cannot be fully used by the latter one.In this article,we devise novel Inner Cascaded U-Net and Inner Cascaded U^(2)-Net as improvements to plain cascaded U-Net for medical image segmentation.The proposed Inner Cascaded U-Net adds inner nested connections between two U-Nets to share more contextual information.To further boost segmentation performance,we propose Inner Cascaded U^(2)-Net,which applies residual U-block to capture more global contextual information from different scales.The proposed models can be trained from scratch in an end-to-end fashion and have been evaluated on Multimodal Brain Tumor Segmentation Challenge(BraTS)2013 and ISBI Liver Tumor Segmentation Challenge(LiTS)dataset in comparison to related U-Net,cascaded U-Net,U-Net++,U^(2)-Net and state-of-the-art methods.Our experiments demonstrate that our proposed Inner Cascaded U-Net and Inner Cascaded U^(2)-Net achieve better segmentation performance in terms of dice similarity coefficient and hausdorff distance as well as get finer outline segmentation.

Design of Network Cascade Structure for Image Super-Resolution

Image super resolution is an important field of computer research.The current mainstream image super-resolution technology is to use deep learning to mine the deeper features of the image,and then use it for image restoration.However,most of these models mentioned above only trained the images in a specific scale and do not consider the relationships between different scales of images.In order to utilize the information of images at different scales,we design a cascade network structure and cascaded super-resolution convolutional neural networks.This network contains three cascaded FSRCNNs.Due to each sub FSRCNN can process a specific scale image,our network can simultaneously exploit three scale images,and can also use the information of three different scales of images.Experiments on multiple datasets confirmed that the proposed network can achieve better performance for image SR.

基于改进Cascade R-CNN的两阶段销钉缺陷检测模型

无人机在输电线路巡检过程中会拍摄大量图片,自动识别无人机拍摄图片中存在的部件缺陷是无人机巡检的重要环节。其中销钉的缺陷由于目标较小且需要依赖上下文信息才能正确判断,识别难度较大。针对上述问题,提出了一种两阶段的销钉缺陷检测模型。首先使用Faster R-CNN(regin convolutional neural networks)模型提取出原始图像中的连接部位,再对提取出的每个连接部位进行缺陷识别。缺陷识别模型使用改进的Cascade R-CNN,该模型使用层级残差卷积模块代替骨干网络中的3×3卷积并使用路径聚合特征金字塔(PAFPN)代替原始网络中的特征金字塔结构,能够有效提取图片中的多尺度特征和上下文信息。最后将级联检测器的最后一级替换为double-head检测器,减少模型误报。实验结果表明,模型对销钉缺失及销钉脱出两类缺陷的平均识别精度能够达到81.2%,与原始的Cascade R-CNN相比提升了7.8%。

基于CascadeR-CNN算法的输电线路小目标缺陷检测方法

输电线路无人机航拍图像缺陷识别是维护线路安全运行的重要巡检手段,但目前的识别算法对于销钉、螺母等小目标缺陷存在识别精确度低、易漏判等问题。将Cascade RCNN算法应用于输电线路缺陷检测中,利用ResNet101网络进行特性提取,增强的网络的特征提取能力,并利用多层级联检测器对输电线路小目标进行判别和分类。基于无人机航拍图像数据集进行实验,实验结果表明,相比于Yolov3检测器和Lighthead R-CNN检测器,Cascade R-CNN算法提高了小目标缺陷检测中的召回率和精确度。

Defect Detection Algorithm of Patterned Fabrics Based on Convolutional Neural Network

The background pattern of patterned fabrics is complex,which has a great interference in the extraction of defect features.Traditional machine vision algorithms rely on artificially designed features,which are greatly affected by background patterns and are difficult to effectively extract flaw features.Therefore,a convolutional neural network(CNN)with automatic feature extraction is proposed.On the basis of the two-stage detection model Faster R-CNN,Resnet-50 is used as the backbone network,and the problem of flaws with extreme aspect ratio is solved by improving the initialization algorithm of the prior frame aspect ratio,and the improved multi-scale model is designed to improve detection of small defects.The cascade R-CNN is introduced to improve the accuracy of defect detection,and the online hard example mining(OHEM)algorithm is used to strengthen the learning of hard samples to reduce the interference of complex backgrounds on the defect detection of patterned fabrics,and construct the focal loss as a loss function to reduce the impact of sample imbalance.In order to verify the effectiveness of the improved algorithm,a defect detection comparison experiment was set up.The experimental results show that the accuracy of the defect detection algorithm of patterned fabrics in this paper can reach 95.7%,and it can accurately locate the defect location and meet the actual needs of the factory.

基于改进Cascade R-CNN网络模型的防振锤缺陷识别

针对高压输电线路中防振锤的背景复杂、缺陷目标小及类别数量不均衡问题,提出一种改进的Cascade R-CNN(cascade region convolutional neural networks)网络模型,用于防振锤的缺陷识别.将SE(squeeze and excitation)模块嵌入ResNet-101(residual network-101),以增强网络学习能力.引入FPN(feature pyramid networks)模块提取多尺度的缺陷特征.利用Focal Loss函数降低Cascade R-CNN候选区域提取模块的分类损失.实验结果表明:相对于其他4种模型,该文模型有相对高的识别准确率;识别防振锤缺陷的效果良好.因此,该文模型具有有效性.

改进YOLOv5的织物缺陷检测方法

为了在不增加网络参数量的条件下提升深度学习方法对织物缺陷检测的精度,提出了一种基于改进YOLOv5的织物缺陷检测方法。通过深度卷积改造通道注意力,剪裁最大池化优化空间注意力,并通过二者构建的双级联注意力机制来搭建特征提取子网络,从而提高网络对缺陷区域纹理和语义特征的提取能力;采用鬼影混洗卷积改进特征融合子网络,强化对提取特征的筛选,在降低模型参数量的同时,改善缺陷信息丢失和无效信息冗余问题;在检测端引入具有角度损失的新型损失函数SIOU,来促进真实框和预测框的拟合并提升对缺陷预测的准确性。实验结果表明:改进的YOLOv5方法在降低YOLOv5基准模型复杂度和计算量的同时,与YOLOv7等六种先进方法相比,可获得更高的检测精度,相较原模型mAP@0.5值提高了2.6个百分点,mAP@0.5:0.9值提高了1.3个百分点。

融合CNN与Transformer的MRI脑肿瘤图像分割

为解决卷积神经网络(convolutional neural network,CNN)在学习全局上下文信息和边缘细节方面受到很大限制的问题,提出一种同时学习局语义信息和局部空间细节的级联神经网络用于脑肿瘤医学图像分割。首先将输入体素分别送入CNN和Transformer分支,在编码阶段结束后,采用一种双分支融合模块将2个分支学习到的特征有效地结合起来以实现全局信息与局部信息的融合。双分支融合模块利用哈达玛积对双分支特征之间的细粒度交互进行建模,同时使用多重注意力机制充分提取特征图通道和空间信息并抑制无效的噪声信息。在BraTS竞赛官网评估了本文方法,在BraTS2019验证集上增强型肿瘤区、全肿瘤区和肿瘤核心区的Dice分数分别为77.92%,89.20%和81.20%。相较于其他先进的三维医学图像分割方法,本文方法表现出了更好的分割性能,为临床医生做出准确的脑肿瘤细胞评估和治疗方案提供了可靠依据。

面向超分辨率重建的层次间局部特征增强网络

基于卷积神经网络的超分辨率重建模型以单项传播为主,层次越靠后感知信息的能力越微弱,导致层次间局部特征部分丢失,难以实质提升网络的特征表达能力。针对此问题,提出层次间局部特征增强网络。该方法由级联残差模块、层次间特征增强块和特征感知注意力机制组成。级联残差模块通过有效残差连接增加对残差分支信息的利用;层次间特征增强块提取不同深度特征的依赖关系,自适应调整中间层特征权值增强捕获关键信息的能力;特征感知注意力机制采用方向感知和位置判断的方式准确定位和识别感兴趣对象。多项标准数据集的实验结果表明,该方法能改善超分辨率的视觉重建效果,整体性能优于现有方法。

基于DCM‑PCA和GA‑BP的逆变器故障诊断

针对光伏并网三相电压型逆变器开关管的开路故障,提出深度级联模型(deep cascade mode,DCM)‒主成分分析(principal component analysis,PCA)与遗传算法(genetic algorithm,GA)优化的BP神经网络结合的故障诊断方法。首先对逆变器的开路故障进行分析和仿真,确定三相电流作为故障信号,选择22类故障状态作为诊断对象,通过以稀疏表示分类(sparse representation based classififier,SRC)为基本操作单元的深度级联模型提取故障特征,DCM根据层次学习特性将故障特征分层,再由SRC部分得到不同故障的编码系数,并采用t分布—随机近邻嵌入(t⁃distributed stochastic neighbor embedding,t⁃SNE)方法验证了DCM具有较好的特征提取能力,通过PCA降低故障特征的冗余度、保留有价值的主成分提高网络映射能力,最后将故障特征向量作为GA⁃BP神经网络的输入信号实现对故障的诊断识别。通过仿真实验得到该方法的故障诊断准确率为95.64%,与DCM⁃PCA⁃BP、FFT⁃GA⁃BP和FFT⁃BP相比准确率分别提高8.71%、20.64%、51.70%,表明该方法有更好的故障特征提取能力和故障诊断效果。

基于区域渐进校准网络的人脸检测与定位

为解决角度变化下的人脸检测中存在参数量大及角度幅度变量小的问题,提出区域渐进校准网络用于任意平面角度的人脸检测,通过级联网络结构降低角度变化、提升网络运行速度。采用区域生成网络产生高质量的候选区域,构造渐进校准网络,逐步缩小面部平面角度变化范围,同时由粗到细地对候选区域执行面部检测。其中,特征提取的中间层融合参数量较少时,更好地表示了面部特征,调整锚的设置解决小尺度面部问题。在角度增强的FDDB(face detection data set and benchmark)数据集与WIDER FACE数据集上的实验结果表明,提出的方法分别取得了89.1%与90.4%的平均召回率,准确度高于快速区域卷积神经网络(Faster RCNN),且运行速度更快。在实际项目中使用该算法,验证了该方法的有效性及可行性。

基于改进的GoogleNet-ResNet算法的路基病害智能分类方法

针对路基病害分类算法存在的复杂病害辨识难度大、多视图雷达图像特征利用不充分等问题,提出一种基于改进的GoogleNet-ResNet算法的路基病害智能分类方法;首先,引入坐标注意力和改进的Inception模块对GoogleNet网络结构进行优化;然后,利用改进的GoogleNet学习c-scan数据特征剔除非目标病害,实现病害目标的粗分类;最后,将分类成病害的b-scan数据输入基于迁移学习的ResNet50,实现病害的细分类;实验表明,改进的GoogleNet进行病害粗分类的准确率可达到98.2%,检测速度可达90.9 fps;基于迁移学习的ResNet50进行病害细分类的准确率可达90.5%,检测速度可达52.6 fps;该算法的准确率比单独的改进的GoogleNet网络高10.1%,比单独的ResNet50网络高7.4%,有效地提高了道路路基病害的识别精度与效率。