Multi-Layer Feature Extraction with Deformable Convolution for Fabric Defect Detection

In the textile industry,the presence of defects on the surface of fabric is an essential factor in determining fabric quality.Therefore,identifying fabric defects forms a crucial part of the fabric production process.Traditional fabric defect detection algorithms can only detect specific materials and specific fabric defect types;in addition,their detection efficiency is low,and their detection results are relatively poor.Deep learning-based methods have many advantages in the field of fabric defect detection,however,such methods are less effective in identifying multiscale fabric defects and defects with complex shapes.Therefore,we propose an effective algorithm,namely multilayer feature extraction combined with deformable convolution(MFDC),for fabric defect detection.In MFDC,multi-layer feature extraction is used to fuse the underlying location features with high-level classification features through a horizontally connected top-down architecture to improve the detection of multi-scale fabric defects.On this basis,a deformable convolution is added to solve the problem of the algorithm’s weak detection ability of irregularly shaped fabric defects.In this approach,Roi Align and Cascade-RCNN are integrated to enhance the adaptability of the algorithm in materials with complex patterned backgrounds.The experimental results show that the MFDC algorithm can achieve good detection results for both multi-scale fabric defects and defects with complex shapes,at the expense of a small increase in detection time.

An improved deep dilated convolutional neural network for seismic facies interpretation

With the successful application and breakthrough of deep learning technology in image segmentation,there has been continuous development in the field of seismic facies interpretation using convolutional neural networks.These intelligent and automated methods significantly reduce manual labor,particularly in the laborious task of manually labeling seismic facies.However,the extensive demand for training data imposes limitations on their wider application.To overcome this challenge,we adopt the UNet architecture as the foundational network structure for seismic facies classification,which has demonstrated effective segmentation results even with small-sample training data.Additionally,we integrate spatial pyramid pooling and dilated convolution modules into the network architecture to enhance the perception of spatial information across a broader range.The seismic facies classification test on the public data from the F3 block verifies the superior performance of our proposed improved network structure in delineating seismic facies boundaries.Comparative analysis against the traditional UNet model reveals that our method achieves more accurate predictive classification results,as evidenced by various evaluation metrics for image segmentation.Obviously,the classification accuracy reaches an impressive 96%.Furthermore,the results of seismic facies classification in the seismic slice dimension provide further confirmation of the superior performance of our proposed method,which accurately defines the range of different seismic facies.This approach holds significant potential for analyzing geological patterns and extracting valuable depositional information.

Fast and Accurate Machine Learning Inverse Lithography Using Physics Based Feature Maps and Specially Designed DCNN

Inverse lithography technology(ILT)is intended to achieve optimal mask design to print a lithography target for a given lithography process.Full chip implementation of rigorous inverse lithography remains a challenging task because of enormous computational resource requirements and long computational time.To achieve full chip ILT solution,attempts have been made by using machine learning techniques based on deep convolution neural network(DCNN).The reported input for such DCNN is the rasterized images of the lithography target;such pure geometrical input requires DCNN to possess considerable number of layers to learn the optical properties of the mask,the nonlinear imaging process,and the rigorous ILT algorithm as well.To alleviate the difficulties,we have proposed the physics based optimal feature vector design for machine learning ILT in our early report.Although physics based feature vector followed by feedforward neural network can provide the solution to machine learning ILT,the feature vector is long and it can consume considerable amount of memory resource in practical implementation.To improve the resource efficiency,we proposed a hybrid approach in this study by combining first few physics based feature maps with a specially designed DCNN structure to learn the rigorous ILT algorithm.Our results show that this approach can make machine learning ILT easy,fast and more accurate.

An Emotion Analysis Method Using Multi-Channel Convolution Neural Network in Social Networks

As an interdisciplinary comprehensive subject involving multidisciplinary knowledge,emotional analysis has become a hot topic in psychology,health medicine and computer science.It has a high comprehensive and practical application value.Emotion research based on the social network is a relatively new topic in the field of psychology and medical health research.The text emotion analysis of college students also has an important research significance for the emotional state of students at a certain time or a certain period,so as to understand their normal state,abnormal state and the reason of state change from the information they wrote.In view of the fact that convolutional neural network cannot make full use of the unique emotional information in sentences,and the need to label a large number of highquality training sets for emotional analysis to improve the accuracy of the model,an emotional analysismodel using the emotional dictionary andmultichannel convolutional neural network is proposed in this paper.Firstly,the input matrix of emotion dictionary is constructed according to the emotion information,and the different feature information of sentences is combined to form different network input channels,so that the model can learn the emotion information of input sentences from various feature representations in the training process.Then,the loss function is reconstructed to realize the semi supervised learning of the network.Finally,experiments are carried on COAE 2014 and self-built data sets.The proposed model can not only extract more semantic information in emotional text,but also learn the hidden emotional information in emotional text.The experimental results show that the proposed emotion analysis model can achieve a better classification performance.Compared with the best benchmark model gram-CNN,the F1 value can be increased by 0.026 in the self-built data set,and it can be increased by 0.032 in the COAE 2014 data set.

Instance Retrieval Using Region of Interest Based CNN Features

Recently, image representations derived by convolutional neural networks(CNN) have achieved promising performance for instance retrieval, and they outperformthe traditional hand-crafted image features. However, most of existing CNN-based featuresare proposed to describe the entire images, and thus they are less robust to backgroundclutter. This paper proposes a region of interest (RoI)-based deep convolutionalrepresentation for instance retrieval. It first detects the region of interests (RoIs) from animage, and then extracts a set of RoI-based CNN features from the fully-connected layerof CNN. The proposed RoI-based CNN feature describes the patterns of the detected RoIs,so that the visual matching can be implemented at image region-level to effectively identifytarget objects from cluttered backgrounds. Moreover, we test the performance of theproposed RoI-based CNN feature, when it is extracted from different convolutional layersor fully-connected layers. Also, we compare the performance of RoI-based CNN featurewith those of the state-of-the-art CNN features on two instance retrieval benchmarks.Experimental results show that the proposed RoI-based CNN feature provides superiorperformance than the state-of-the-art CNN features for in-stance retrieval.

无人机双目视觉鲁棒定位方法

无人机(unmanned aerial vehicle,UAV)在全球定位系统(global positioning system,GPS)信号拒止环境中的应用受到限制,传统视觉同步定位与建图(simultaneous localization and mapping,SLAM)技术一定程度上解决了该问题,但在动态场景和弱纹理场景中定位精度较差。针对该问题提出一种基于双目视觉的多场景鲁棒SLAM方法,重点考虑了真实环境中的动态和弱纹理2类具有挑战性的场景,利用双目相机为UAV在动态和弱纹理场景中提供位姿信息。针对动态场景利用掩膜基于区域的卷积神经网络(mask region-based convolutional neural network,Mask R-CNN)分割潜在动态内容并剔除动态特征,通过计算稠密光流同步相邻帧的掩膜,减小了掩膜的计算成本。对于弱纹理场景,在传统SLAM算法使用的点特征基础上融合了线特征,充分利用了环境中的结构特征。数值模拟和仿真实验证明了本文算法具有更高的鲁棒性和精确性。

基于锚点的快速三维手部关键点检测算法

在人机协作任务中,手部关键点检测为机械臂提供目标点坐标,A2J(Anchor-to-Joint)是具有代表性的一种利用锚点进行关键点检测的方法。A2J以深度图为输入,可实现较好的检测效果,但对全局特征获取能力不足。文中设计了全局-局部特征融合模块(Global-Local Feature Fusion,GLFF)对骨干网络浅层和深层的特征进行融合。为了提升检测速度,文中将A2J的骨干网络替换为ShuffleNetv2并对其进行改造,用5×5深度可分离卷积替换3×3深度可分离卷积,增大感受野,有效提升了骨干网络对全局特征的提取能力。文中在锚点权重估计分支引入高效通道注意力模块(Efficient Channel Attention,ECA),提升了网络对重要锚点的关注度。在主流数据集ICVL和NYU上进行的训练和测试结果表明,相比于A2J,文中所提方法的平均误差分别降低了0.09 mm和0.15 mm。在GTX1080Ti显卡上实现了151 frame·s^(-1)的检测速率,满足人机协作任务对于实时性的要求。

基于双重注意力机制的人群计数方法

针对复杂场景下人群计数问题中的尺度变化、背景干扰和部分遮挡等问题,在空洞卷积操作的基础上,提出一种基于双重注意力机制的空洞上下文卷积神经网络(DA-DCCNN)。首先,将VGG16中的卷积层作为特征提取器,获取人群图像抽象、深层的特征图;其次,利用空洞卷积构造空洞上下文模块(DCM)对不同层获取的特征进行连接,并引入空间注意力模块(SAM)和通道注意力模块(CAM)获取上下文信息;最后,组合欧氏距离和交叉熵构造损失函数,对网络预测注意力图和真实注意力图之间的差异进行度量。在ShanghaiTech、UCF_CC_50和UCF-QNRF 3个公开数据集上的实验结果表明,DA-DCCNN在有效获取图像的多尺度特征的同时,增强了对图像中重要区域和通道的感知能力,平均绝对误差(MAE)取得了相对最优的结果。基于双重注意力机制的特征融合网络能有效感知图像中的空间结构和局部特征,从而使得生成的密度图能更准确地对人群区域进行预测和计数。

基于轻量级CNN的视觉SLAM快速回环检测算法

传统基于卷积神经网络(Convolutional Neural Network,CNN)的视觉同步定位与地图构建(Simultaneous Localization and Mapping,SLAM)系统回环检测目前准确率和召回率较高,但其存在特征提取时间较长和特征向量维度过高导致计算量较大等问题。针对上述问题,结合轻量级卷积神经网络MobileNetV3和PCA降维算法,提出了一种基于深度学习的快速回环检测算法。基于MobileNetV3进行特征提取并构建特征矩阵,运用PCA降维算法完成降维以提升运行速度,使用余弦相似度计算各个特征向量间的相似性,并取最大值与给定阈值比较判断是否构成回环。最后,使用New College和City Centre两个公开的数据集验证算法的性能。实验结果表明,相较于传统的CNN回环检测方法,提出的算法在保证准确率和召回率的同时,运行速度更快,较好的满足了视觉SLAM系统准确性和实时性的要求。

基于多分支特征融合的密集人群计数网络

针对人群计数任务中存在的多尺度变化、背景噪声等问题,提出一种基于多分支特征融合的人群计数网络。在网络前端设计一个双向特征融合路径,将网络深层的语义信息和浅层的空间细节信息进行反复提取融合,使用位置注意力机制和通道注意力机制增强网络对人群和背景之间的判别能力,生成高质量特征图;网络后端采用密集残差连接增强网络对人头连续的多尺度信息提取能力,得到最终的人群密度图。在ShanghaiTech、UCF_CC_50和UCF_QNRF数据集上分别进行的对比实验的结果表明,该模型的计数性能优于先前诸多方法,有着良好的计数精度。

基于多尺度特征聚合和密集连接的人群计数网络

人群计数任务处理的图像受到遮挡、视角变化和透视效应、背景成像干扰等多方面的挑战。针对复杂场景中人群计数任务存在的尺度变化、图片噪声等问题,设计一种基于多尺度特征聚合和密集连接的人群计数网络模型。模型的一个重要组成部分是密集连接而成的多尺度特征聚合模块,它通过不同卷积核提取多尺度特征并聚合其跨尺度的信息进行更准确的估计。该网络模型在三个公开数据集Shanghai Tech、UCF_QNRF、UCF_CC_50上进行测试,实验结果表明,该模型相比目前先进算法CSRNet在平均绝对误差(MAE)、均方误差(MSE)均有不同程度的减少,模型精度更好。与其他模型相比,该模型更充分地利用多尺度特征信息,提高了人群计数任务的精度。

基于多尺度特征融合的地理测绘影像目标检测

为了提高对地理测绘目标的检测准确度,设计了基于多尺度特征融合的地理测绘影像目标检测方法。初步提取地理测绘遥感影像的边缘信息,并计算其边缘密度与边缘分布情况,通过增强边缘信息实现对遥感影像的预处理,得到更明确的影像边缘信息;利用梯度采样法建立下降金字塔影像,并融合多尺度特征,为后续的目标提取提供更准确、特征更明显的信息;根据特征融合结果,采用深度卷积网络实现对地理测绘影像目标的有效检测。结果表明,应用该方法,检测结果的准确率、召回率和F 1分数数值均较高,检测耗时也维持在较低的数值范围,该方法可明显提高目标检测效果。

基于卷积神经网络的无人机遥感测绘图像特征提取方法

在地质勘探学领域,无人机遥感技术已成为获取地表数据的重要手段,而卷积神经网络(CNN)因其出色的图像处理能力被广泛应用于图像特征提取。文中介绍了一种基于卷积神经网络的无人机遥感测绘图像特征提取方法,分析了CNN的基本概念与结构、无人机遥感图像的特征类型、网络结构设计、数据预处理、特征提取及特征融合过程,旨在提高地质勘探中无人机遥感图像分析的自动化和准确性。

融合注意力特征及动态卷积的肺结节辅助诊断

针对肺结节关键影像征象信息不易获取,部分卷积神经网络(convolutional neural networks,CNN)模型对肺结节的识别率不高的问题,提出一种融合注意力特征的动态卷积残差网络(dynamic convolutional residual networks incorporating attention features,DcANet),并在有效实现肺结节良恶性分类的基础上对所提模型的诊断结果进行可视化分析。此网络以适应肺结节三维小尺寸输入特点的残差网络为基本框架,在DcABlock部分使用可以自适应调整卷积参数的动态卷积以及迭代注意特征融合模块,使模型能够更准确地获取肺结节信息,提高模型的表征能力。此外,还使用类激活映射将三维图像的各层切片进行可视化分析。实验在最终测试集上的准确率为85.87%,平衡F分数(F1)值为82.67%,敏感度和特异性的综合指标Gmean值为85.51%。实验结果表明:该网络可以提升对肺结节良恶性分类的准确性,诊断结果具有可信性,有一定的临床应用价值。

基于特征融合的牵引电机轴承声学故障诊断

滚动轴承作为高速列车牵引电机的重要部件,其故障情况严重影响列车运行安全。声学轴承故障诊断方式具有无安装侵入性、运维成本低的优点,但也具有信噪比低、故障特征难以提取的缺点,机器学习则具有克服噪声影响的鲁棒性。针对应用机器学习进行声学故障诊断时,少量特征无法全面表征轴承故障的难题,文章提出将格拉姆角场(GAF)与小波时频图进行叠加融合,构成6通道融合特征图用以有效表征轴承的故障。首先,建立牵引电机轴承声学故障试验台获取故障声学信号;其次,建立基于GAF的声学信号融合特征图,然后使用残差网络(ResNET)模型针对融合特征图特征训练并验证故障分类模型,并与以单种特征图作为特征的故障分类方法进行准确率对比。结果表明,基于GAF的融合特征图的声学故障分类模型具有99.89%的准确率,融合特征图能更有效地映射轴承故障。

基于特征图融合的对抗样本生成方法

为检验现有深度学习算法的鲁棒性和安全性,提出一种基于特征图融合的对抗样本生成方法。首先,分析卷积神经网络在图像分类任务中所提取不同层次特征图的特点,提出利用多层次特征图进行对抗扰动构造的方法思想;其次,引入通道注意力模块对卷积层输出特征图进行权重分配,以此代表不同特征图对分类结果的影响程度;再次,构建基础网络用于选取高权重特征图,并对显著特征信息进行像素值修改来生成扰动特征图;最后,将不同扰动特征图融合为对抗扰动,并添加至原始输入样本中生成对抗样本。实验结果表明,所提对抗样本生成方法在CIFAR-10和MNIST数据集上兼顾了攻击成功率和样本视觉感知效果,与现有代表性对抗样本生成方法相比,在高难度的非交互式黑盒模型上取得了较好的攻击效果。

基于多传感器融合声纹特征图谱的变压器铁芯松动故障诊断方法

变压器铁芯轻微松动故障给变压器安全稳定运行留下巨大隐患,目前尚缺乏切实可靠的诊断方法。提出一种基于多传感器融合声纹特征图谱的变压器铁芯松动故障诊断方法。首先,利用4个传感器采集声纹时序序列,通过小波变换生成声纹特征图谱,利用熵权法确定不同传感器信号的权重分配,将4个声纹特征图谱加权融合,从而形成多传感器融合声纹特征图谱。其次,将融合声纹特征图谱输入优化后的ShuffleNetV2模型,通过分组卷积和通道混洗得到铁芯松动程度。最后,通过现场试验验证了方法的有效性。结果表明,所提方法对25%,50%,75%及100%的松动程度均能实现可靠诊断,平均准确率高达99.6%。与采用傅里叶频谱(fast Fourier transform, FFT)、格拉米角场(Gramian angular field, GAF)、马尔可夫变换场(Markov transform field, MTF)以及混沌特征(recurrence plot, RP)等传统声纹特征图谱的诊断相比,所提方法识别准确率提高了12.2%;与采用单传感器声纹特征图谱的诊断相比,所提方法识别准确度提高了5.8%;与采用AlexNet、MobilleNetV2、GoogleNet以及ResNet等卷积神经网络模型的诊断相比,所提方法识别准确率提高了2.7%。

基于特征融合编解码的人群计数和密度估计

为解决人群计数任务中因人群大小不一导致的计数偏差问题,提出一种基于特征融合的编解码卷积神经网络模型(CFFNet)。前端网络模块对输入的人群图像自动编码,提取不同尺度的人群特征语义信息;后端网络模块对编码后的人群特征信息进行融合和解码,得到最终的估计密度图。将该模型在4个公开数据集上进行实验,并与历年的主要方法进行对比,实验结果表明,该模型在ShanghaiTech PartA、UCSD和Mall数据集上取得了更好的MAE指标,优于目前的这些算法,验证了模型对不同的人群尺度具有很好的适应性。

跨模态融合和边界可变形卷积引导的RGB-D显著性目标检测

RGB-Depth(RGB-D)显著性目标检测是一项有意义且具有挑战性的任务,基于现有卷积神经网络检测方法在简单场景中获得了良好的检测性能,但不能有效应对背景信息混乱,深度图质量低和目标轮廓复杂的情况.为应对上述问题,本文提出了一种跨模态融合和边界可变形卷积引导的RGB-D显著性目标检测方法.首先,本文以Swin-Transformer为特征提取器,分别对RGB模态与深度图模态进行特征提取,并通过跨模态注意力增强特征模块对两种模态特征进行融合以挖掘显著物的共性与互补特征.接着将提出的相邻多尺度特征增强模块嵌入编码器深层,以获得丰富的全局上下文特征信息,更精准地定位显著物的位置.然后通过构建一个边界特征提取解码器(U-Net架构)生成显著物的边界线索图,并重复采用跨模态融合特征确保生成显著物边界的完整性.最后,本文设计了一个边界可变形卷积引导模块,使用边界线索图与可变形卷积引导跨模态融合特征进行解码以得到更加准确的显著图.通过在6个公开基准数据集上与25种主流方法相比较,本文所提模型在多个指标上均有较明显的提升,从而证明了本文方法的有效性.

基于三维特征图和改进DenseNet的脑电情绪识别方法

情感作为人脑的高级功能,对人的心理健康状态有很大的影响。为了充分考虑脑电信号的空间信息以及时频信息,更好地实现人机交互,论文提出了1种基于三维特征图的改进DenseNet情绪识别模型。通过提取脑电信号θ、α、β和γ 4个频段的微分熵特征,结合脑电通道电极的位置映射关系,构造三维特征图,最后使用改进DenseNet网络进行二次特征提取与分类。为了验证该方法的有效性,在SEED数据集上进行了包含积极、中性、消极3种情绪的分类实验,单被试者实验和所有被试者实验获得的分类准确率分别达98.51%和98.68%。实验结果表明,三维特征图结合特征重用方法能够得到高精度的分类结果,为情绪识别提供了可以尝试的新方向。