基于迁移学习的钢金相组织分类与识别方法的研究

金相检验是分析钢内部组织的常用方法,其中检验图像由人工判别,容易受到主观因素的影响而造成结果的不确定.近年来,深度学习(Deep learning,DL)方法中的卷积神经网络(Convolutional neural networks,CNN)能从原始图像中学习复杂的特征,在图像分类与识别领域得到了广泛的应用.CNN建模需要大量的训练样本才能达到较好的泛化能力,材料科学与工程领域针对具体问题的数据集往往较小,不能满足CNN建模的条件,制约了DL在材料领域的应用.本研究基于lmageNet数据集预训练VGG19模型,对火力发电机组耐热钢金相检验图像进行识别,采用冻结全部卷积层权值和微调部分卷积层权值两种迁移学习方法,可以克服金相图像数据集较小的问题,实现小样本数据集的深度学习建模,两种方法的准确率分别为92.5％和94.2％.微调方式的迁移学习CNN模型具有较快的收敛速度、较高的训练精度与泛化能力,能够对金相组织图像进行较为准确的分类与识别,是一种智能的钢金相组织识别方法,也是自动化分类与识别钢金相组织的一种新方法.     

基于卷积神经网络的典型农作物叶病害识别算法

目前,典型的一些农作物叶病害诊断主要依靠人工,但该方式耗时费力。针对大豆、棉花、水稻、小麦和玉米5类典型农作物的常见叶病害诊断问题,提出了一种基于卷积神经网络的典型农作物叶病害识别方法。从Plantvillage数据库以及其他网站收集典型农作物的叶病害图像,并对这些图像进行预处理,构建了含12 836张的数据集。参照AlexNet框架,构建8层卷积神经网络,采用迁移学习训练网络,最后通过测试集验证网络的识别准确率和损失值。分析不同的卷积神经网络的性能,实验结果表明：本算法对典型农作物的叶病害有良好的识别效果;迁移学习模式下,学习率为0.001时本算法在训练集的识别准确率约为99.47%,在测试集的识别准确率约为96.18%。     

重参数化VGG网络在滚动轴承故障诊断中的应用研究

基于神经网络的滚动轴承故障诊断方法训练时,存在诊断准确率低和易受到变工况噪声干扰的问题,提出一种基于重参数化VGG(RepVGG)滚动轴承故障诊断方法。为满足神经网络对数据量的要求,采用数据增强技术来扩充原始数据,使用短时傅里叶变换(STFT)对原始的振动信号处理成单通道时频图,并使用伪彩色处理技术转换成三通道时频图,进一步将数据输入到RepVGG网络的不同结构中进行滚动轴承的故障诊断。在凯斯西储大学(CWRU)滚动轴承数据集上开展试验验证,试验结果表明,RepVGG在变工况及噪声干扰下的平均诊断准确率分别为98.02%、95%以上,高于基于VGG、ResNet的故障诊断模型,有较高的故障诊断准确率且泛化性更好。     

基于注意力机制改进残差神经网络的轴承故障诊断方法EI北大核心CSCD

针对滚动轴承在不同工况环境中故障诊断训练时间长、准确率低和泛化性能弱的问题,提出了基于注意力机制改进残差神经网络的轴承故障诊断方法。为了提高ResNet模型的准确率和泛化性,提出了基于注意力机制的SE-ResNet模型和CBAM-ResNet模型,并在凯斯西储大学数据集上进行了试验,在同工况有训练集的情况下ResNet模型测试的准确率为97.28%,在不同工况下模型直接迁移的准确率为94.14%~96.86%,CBAM-ResNet模型在不同工况下模型直接迁移的准确率为97.14%~98.86%,SE-ResNet模型在不同工况下模型直接迁移的准确率为97.86%~99.71%,两种改进模型的准确率都明显优于原ResNet模型,表明提出的优化模型提高了ResNet模型的准确率和泛化性。     

基于一维卷积注意力门控循环网络和迁移学习的轴承故障诊断EI北大核心CSCD

针对实际应用中轴承故障数据获取困难,导致其故障诊断模型的准确性和泛化能力下降的问题,提出了一种基于一维卷积注意力门控循环网络(1DCNN-Attention-GRU)和迁移学习的轴承故障诊断方法。首先,构建基于一维卷积网络(1D convolutional neural networks, 1DCNN)、门控循环网络(gated recurrent unit, GRU)和注意力机制(Attention)的故障诊断网络,解决传统故障诊断方法特征提取过于依赖人工经验的问题;然后,引入迁移学习,利用充足的源域数据训练网络,冻结训练好的网络底层结构,再利用少量目标域数据对网络顶层结构进行微调,得到目标网络模型;最后,采用Softmax函数进行故障分类。实验结果表明,所提方法在不同训练样本比例下的故障诊断准确率均高于1DCNN-GRU、GRU以及支持向量机(support vector machine, SVM),并能在变工况及小样本数据下获得较高的故障诊断准确率,当使用3%目标域数据进行微调时,故障诊断准确率高于98%。     

基于卷积神经网络的PCB缺陷图像识别

PCB缺陷图像检测是确保PCB生产质量的重要环节,但传统的人工PCB缺陷检测具有劳动强度大、工作效率低等不足.为此,本文研究了一种基于卷积神经网络的PCB缺陷图像识别方法,建立了包括三种PCB缺陷和无缺陷图像的数据集,基于ResNet101网络模型搭建了PCB缺陷图像识别分类模型.引入迁移学习的方法,基于在大数据集上充分训练好的模型结合PCB图像数据集,并训练该PCB缺陷图像识别模型.实验结果表明,ResNet101模型对无缺陷PCB图像和三类常见PCB缺陷图像的平均识别准确率达到91.98％,验证了该模型对PCB图像识别分类的有效性.     

基于迁移学习的钢丝绳断丝定量检测方法EI北大核心CSCD

针对目前钢丝绳断丝定量检测的问题,利用深度卷积神经网络强大的特征提取能力,提出一种基于迁移学习的钢丝绳断丝定量识别方法。通过连续小波变换将原始断丝漏磁信号转换成时频图。将预训练网络GoogLeNet的低层参数直接迁移,使用标记好的时频图对网络高层进行参数调整,得到最终的目标模型。通过内外部断丝试验验证了所提出的定量识别模型的效果,将传统的BP(back propagation)神经网络与所提出的方法进行对比。结果表明:基于迁移学习的断丝定量识别方法能准确区分钢丝绳的内外部断丝故障,分类准确率达到了97.2%;与传统BP神经网络相比,所提出的方法对各种断丝具有更好的识别性能。     

基于伪标签深度学习的半监督滚动轴承故障诊断模型

针对实际工程应用中被标记的滚动轴承故障样本收集困难,传统诊断模型精度较低的问题,提出一种伪标签学习融合参数迁移深度学习网络的半监督滚动轴承故障诊断模型。首先将ImageNet数据集上预训练的残差网络（Residual Network,ResNet）模型参数迁移至本文模型中作为初始参数,并使用不同学习率微调网络层参数以加快模型收敛速度;随后引入伪标签半监督学习,使用标签数据训练模型并对无标记数据进行预测以生成伪标签;最后使用标签数据以及伪标签数据训练参数迁移后的ResNet模型,并测试诊断效果。对两种滚动轴承故障数据进行半监督下故障诊断实验及跨域故障诊断实验。实验结果表明,在具有大量未标记样本集下,所提出模型可迁移至不同设备完成诊断,具有较强的鲁棒性,可用于处理复杂工业环境中的故障诊断问题。     

基于VGG16卷积神经网络的5G高噪调制识别技术

随着第五代移动通信技术（5th Generation Mobile Communication Technology,5G）的快速发展,电磁信道环境趋向复杂。以采用人工智能技术提高5G通信调制识别率为目的,在低信噪比的高斯噪声背景下,模拟适用于神经网络算法模型输入的5G星座图数据集,基于Tensorflow2.0平台搭建16层几何视觉群网络（Visual Geometry Group Network-16 Layers,VGG16）卷积神经网络模型,通过算法设计与仿真试验,验证了高噪环境下5G调制识别技术的可行性。与业界现有的传统调制识别技术相比,采用VGG16卷积神经网络对5G的调制方式进行识别,可大幅提高识别的准确率,在实际的工程应用中取得了很好的效果。     

基于深度学习的水体色度分类算法研究

运用深度学习技术进行非接触、快速水体色度检测与分类,采用无人机采集水体图像,运用色度仪对标定的图像完成分类,建立数据集。采用图像归一化处理减少环境因素对分类结果的影响,设计多特征的分步边缘检测算法,检测水域图像边缘,剔除无关像素。对VGG 16、GoogleNet-V3和ResNet 18卷积神经网络进行水体色度分类模型构建与训练,后筛选Inception结构和残差结构为基本构建单元,设计专门用于水体色度分类的WCNet 15与WCNet 21神经网络模型。在训练集上训练参数并利用验证集完成对2个模型的准确率的比较,筛选准确率高的WCNet 21模型作为最终水体色度分类模型。WCNet21模型的最优准确率可达97.8%,满足水体色度分类需求,可应用到具体的水体色度分类工作当中。     

基于环形特征的卷积神经网络轮毂识别

针对实际生产中不同种类轮毂的混流生产问题,提出了一种基于环形特征的卷积神经网络轮毂识别算法。将直角坐标下的环形轮毂映射到极坐标中,归一化为标准形式的矩形,提取轮毂图像的环形特征信息,减少冗余特征产生的影响;设计了一种改进的VGG网络架构,利用深度可分离卷积打破输出通道维度与卷积核大小的联系,在不损失网络性能的同时降低了计算量,能够在实际生产中轮毂识别任务在有限的算力情况下实时进行计算;从有效性和实时性两个方面对轮毂识别算法进行评估,且通过Inception V3、SVM、KNN等模型的对比实验,验证了该算法可以实时地对轮毂自适应分类。实验表明: 该方法对轮毂图像的处理精度达到99%以上,单幅图像平均处理时间降低至11.78ms。     

Improved VGG Model for Road Traffic Sign Recognition

Road traffic sign recognition is an important task in intelligent transportation system. Convolutional neural networks (CNNs) have achieved a breakthrough in computer vision tasks and made great success in traffic sign classification. In this paper, it presents a road traffic sign recognition algorithm based on a convolutional neural network. In natural scenes, traffic signs are disturbed by factors such as illumination, occlusion, missing and deformation, and the accuracy of recognition decreases, this paper proposes a model called Improved VGG (IVGG) inspired by VGG model. The IVGG model includes 9 layers, compared with the original VGG model, it is added max-pooling operation and dropout operation after multiple convolutional layers, to catch the main features and save the training time. The paper proposes the method which adds dropout and Batch Normalization (BN) operations after each fully-connected layer, to further accelerate the model convergence, and then it can get better classification effect. It uses the German Traffic Sign Recognition Benchmark (GTSRB) dataset in the experiment. The IVGG model enhances the recognition rate of traffic signs and robustness by using the data augmentation and transfer learning, and the spent time is also reduced greatly.     

卷积神经网络在印刷品缺陷检测的应用

目的 为了改善传统机器检测印刷产品缺陷存在误费率高的不足。方法 提出以卷积神经网络为控制核心的印刷品缺陷检测系统。设计可在实际检测中应用的卷积神经网络,设计在线印刷质量检测系统的硬件结构。结果 对结构相同而训练次数、学习率不同的卷积神经网络进行了缺陷检测的性能对比,验证了该卷积神经网络在学习率小于0.01时,可以获得较好的识别效果;在学习率大于0.05时,网络不容易收敛。网络训练次数越多,精度越高,相应的训练时间也较长。在满足快速性和精确度的条件下,确定了适应某印刷品的缺陷检验网络训练次数为50,学习率为0.005,此时的识别率为90%。结论 经过实验证明,该检测系统具有良好的缺陷识别能力,缺陷类型的分类准确率较高。该系统具有一定的实用价值。     

基于一维卷积神经网络的房颤智能诊断方法研究

针对“大数据”时代如何利用数据对房颤进行智能、高效的诊断问题,提出了基于一维卷积神经网络的智能诊断方法,以避免传统算法依赖人工特征提取和先验知识的问题。首先,分别构建一维LeNet-5和AlexNet神经网络模型,合理设置网络结构参数;然后,在采集的实验数据基础上针对心电信号的特点进行一系列的数据处理,随机构建训练样本和测试样本;最后,将训练样本分别输入上述2个神经网络模型中训练学习,再将训练好的模型用于房颤的诊断。实验结果表明:一维LeNet-5网络模型存在“过拟合”现象,而一维AlexNet网络模型在避免了上述现象的同时,诊断精度达到了95.34%,较传统方法有了较大提升,为房颤诊断提供了有效的手段。     

Abnormal Behavior Detection and Recognition Method Based on Improved ResNet Model

The core technology in an intelligent video surveillance system is that detecting and recognizing abnormal behaviors timely and accurately. The key breakthrough point in recognizing abnormal behaviors is how to obtain the effective features of the picture, so as to solve the problem of recognizing them. In response to this difficulty, this paper introduces an adjustable jump link coefficients model based on the residual network. The effective coefficients for each layer of the network can be set after using this model to further improving the recognition accuracy of abnormal behavior. A convolution kernel of 1×1 size is added to reduce the number of parameters for the purpose of improving the speed of the model in this paper. In order to reduce the noise of the data edge, and at the same time, improve the accuracy of the data and speed up the training, a BN (Batch Normalization) layer is added before the activation function in this network. This paper trains this network model on the public ImageNet dataset, and then uses the transfer learning method to recognize these abnormal behaviors of human in the UTI behavior dataset processed by the YOLO_v3 target detection network. Under the same experimental conditions, compared with the original ResNet-50 model, the improved model in this paper has a 2.8% higher accuracy in recognition of abnormal behaviors on the public UTI dataset.     

基于深度学习的易拉罐缺陷检测技术

目的 现有的易拉罐缺陷检测系统在高速生产线中存在错检率和漏检率高,检测精度相对较低等问题,为了提高易拉罐缺陷识别的准确性,使易拉罐生产线实现进一步自动化、智能化,基于深度学习技术和迁移学习技术,提出一种适用于易拉罐制造的在线检测的算法。方法 利用深度卷积网络提取易拉罐缺陷特征,通过优化卷积核,减短易拉罐缺陷检测的时间。针对国内外数据集缺乏食品包装制造的缺陷图像,构建易拉罐缺陷数据集,结合预训练网络,通过调整VGG16提升对易拉罐缺陷的识别准确率。结果 对易拉罐数据集在卷积神经网络、迁移学习和调整后的预训练网络进行了易拉罐缺陷检测的性能对比,验证了基于深度学习的易拉罐缺陷检测技术在学习率为0.0005,训练10个迭代后可达到较好的识别效果,最终二分类缺陷识别率为99.7%,算法耗时119 ms。结论 相较于现有的易拉罐检测算法,文中提出的基于深度学习的易拉罐检测算法的识别性能更优,智能化程度更高。同时,该研究有助于制罐企业利用深度学习等AI技术促进智能化生产,减少人力成本,符合国家制造业产业升级的策略,具有一定的实际意义。     

Residual learning based CNN for breast cancer histopathological image classification

Biopsy is one of the most commonly used modality to identify breast cancer in women, where tissue is removed and studied by the pathologist under the microscope to look for abnormalities in tissue. This technique can be time-consuming, error-prone, and provides variable results depending on the expertise level of the pathologist. An automated and efficient approach not only aids in the diagnosis of breast cancer but also reduces human effort. In this paper, we develop an automated approach for the diagnosis of breast cancer tumors using histopathological images. In the proposed approach, we design a residual learning-based 152-layered convolutional neural network, named as ResHist for breast cancer histopathological image classification. ResHist model learns rich and discriminative features from the histopathological images and classifies histopathological images into benign and malignant classes. In addition, to enhance the performance of the developed model, we design a data augmentation technique, which is based on stain normalization, image patches generation, and affine transformation. The performance of the proposed approach is evaluated on publicly available BreaKHis dataset. The proposed ResHist model achieves an accuracy of 84.34% and an F1-score of 90.49% for the classification of histopathological images. Also, this approach achieves an accuracy of 92.52% and F1-score of 93.45% when data augmentation is employed. The proposed approach outperforms the existing methodologies in the classification of benign and malignant histopathological images. Furthermore, our experimental results demonstrate the superiority of our approach over the pre-trained networks, namely AlexNet, VGG16, VGG19, GoogleNet, Inception-v3, ResNet50, and ResNet152 for the classification of histopathological images.     

Strategy for Rapid Diabetic Retinopathy Exposure Based on Enhanced Feature Extraction Processing

In the modern world, one of the most severe eye infections brought on by diabetes is known as diabetic retinopathy (DR), which will result in retinal damage, and, thus, lead to blindness. Diabetic retinopathy (DR) can be well treated with early diagnosis. Retinal fundus images of humans are used to screen for lesions in the retina. However, detecting DR in the early stages is challenging due to the minimal symptoms. Furthermore, the occurrence of diseases linked to vascular anomalies brought on by DR aids in diagnosing the condition. Nevertheless, the resources required for manually identifying the lesions are high. Similarly, training for Convolutional Neural Networks (CNN) is more time-consuming. This proposed research aims to improve diabetic retinopathy diagnosis by developing an enhanced deep learning model (EDLM) for timely DR identification that is potentially more accurate than existing CNN-based models. The proposed model will detect various lesions from retinal images in the early stages. First, characteristics are retrieved from the retinal fundus picture and put into the EDLM for classification. For dimensionality reduction, EDLM is used. Additionally, the classification and feature extraction processes are optimized using the stochastic gradient descent (SGD) optimizer. The EDLM’s effectiveness is assessed on the KAGGLE dataset with 3459 retinal images, and results are compared over VGG16, VGG19, RESNET18, RESNET34, and RESNET50. Experimental results show that the EDLM achieves higher average sensitivity by 8.28% for VGG16, by 7.03% for VGG19, by 5.58% for ResNet18, by 4.26% for ResNet 34, and by 2.04% for ResNet 50, respectively.     

基于BP神经网络的压电陶瓷蠕变预测

压电陶瓷驱动器的蠕变误差随时间呈现非线性变化,难以实时修正。提出基于BP神经网络的压电陶瓷蠕变预测方法,使用压电陶瓷驱动系统采集数据,对数据进行归一化处理,通过实验设计BP神经网络的隐含层数、隐含层节点数、节点转移函数和训练函数,构建BP神经网络预测模型,建立压电陶瓷蠕变与时间的关系。用BP神经网络模型对压电陶瓷蠕变进行了预测仿真,并将结果与实测数据进行了对比。结果表明,蠕变预测结果与实验数据的最大绝对误差均小于0.1 μm,最大蠕变误差均不超过0.6%,最大均方误差仅为0.0021,可见,BP预测模型具有较高的预测精度,可作为预测压电陶瓷蠕变误差的一种有效手段。