基于深度卷积神经网络的车型识别

传统的基于卷积神经网络的车型识别算法存在识别相似车型的准确率不高,以及在网络训练时只能使用图像的灰度图从而丢失了图像的颜色信息等缺陷。对此,提出一种基于深度卷积神经网络(Deep Convolution Neural Network,DCNN)的提取图像特征的方法,运用深度卷积神经网络对背景较复杂的车型进行网络训练,以达到识别车型的目的。文中采用先进的深度学习框架Caffe,基于AlexNet结构提出了深度卷积神经网络的模型,分别对车型的图像进行训练,并与传统CNN算法进行比较。实验结果显示,DCNN网络模型的准确率达到了96.9%,比其他算法的准确率更高。     

Bayesian neural networks at scale: a performance analysis and pruning study

Bayesian neural networks (BNNs) are a promising method of obtaining statistical uncertainties for neural network predictions but with a higher computational overhead which can limit their practical usage. This work explores the use of high-performance computing with distributed training to address the challenges of training BNNs at scale. We present a performance and scalability comparison of training the VGG-16 and Resnet-18 models on a Cray-XC40 cluster. We demonstrate that network pruning can speed up inference without accuracy loss and provide an open-source software package, BPrune, to automate this pruning. For certain models we find that pruning up to 80% of the network results in only a 7.0% loss in accuracy. With the development of new hardware accelerators for deep learning, BNNs are of considerable interest for benchmarking performance. This analysis of training a BNN at scale outlines the limitations and benefits compared to a conventional neural network.

     

基于降噪自编码神经网络的事件相关电位脑电信号分析方法

提出一种基于降噪自编码神经网络事件相关电位分析方法,首先建立3层神经网络结构,利用降噪自编码对神经网络进行初始化,实现了降噪自编码深度学习模型的无监督学习.从无标签数据中自动学习数据特征,通过优化模型训练得到的权值作为神经网络初始化参数.其次,经过有标签的样本进行网络参数的微调即可完成对神经网络的训练,该方法有效解决了神经网络训练中因随机选择初始化参数,而导致网络易陷入局部极小的缺陷.最后,利用上述神经网络对第3届脑机接口竞赛数据集Data set Ⅱ(事件相关电位脑电信号)进行分类分析.实验结果表明:利用降噪自编码迭代2500次训练神经网络模型,在受试者A和受试者B样本数据叠加5次、10次、15次3种情况下获得的分类准确率分别为73.4%, 87.4%和97.2%.该最高准确率优于其他分类方法,比竞赛第1名联合支持向量机(SVM)分类器(ESVM)提高了0.7%,为事件相关电位脑电信号提供了一种深度学习分析方法.     

基于卷积神经网络和迁移学习的肺结节检测

为解决训练样本不足的问题,提出一种基于卷积神经网络和迁移学习的X光胸片肺结节检测方法。基于Keras深度学习框架,对比分析3种预训练卷积神经网络模型的分类性能,在此基础上进一步探究迁移学习的有效性。在公开的JSRT数据集上进行验证,提出方法获得了93.75%的准确度、94.36%的敏感度、92.74%的特异度以及98.20%的AUC值。与已有的其它研究进行对比,实现了最高的敏感度和较低的假阳性率,验证了迁移学习的有效性和所提算法的可行性。     

Autonomous Learning System Towards Mobile Intelligence

How to efficiently deploy machine learning models on mobile devices has drawn a lot of attention in both academia and industries, among which the model training is a critical part. However, with increasingly public attention on data privacy and the recently adopted laws and regulations, it becomes harder for developers to collect training data from users and thus they cannot train high-quality models. Researchers have been exploring approaches to training neural networks on decentralized data. Those efforts will be summarized and their limitations be pointed out. To this end, this work presents a novel neural network training paradigm on mobile devices, which distributes all training computations associated with private data on local devices and requires no data to be uploaded in any form. Such training paradigm is named autonomous learning. To deal with two main challenges of autonomous learning, i.e., limited data volume and insufficient computing power available on mobile devices, this paper designs and implements the first autonomous learning system AutLearn. It incorporates the cloud (public data and pre-training)--client (private data and transfer learning) cooperation methodology and data augmentation techniques to ensure the model convergence on mobile devices. Furthermore, by optimization techniques such as model compression, neural network compiler, and runtime cache reuse, AutLearn can significantly reduce the on-client training cost. Two classical scenarios of autonomous learning are implemented based on AutLearn,with a set of experiments carried out. The results show that AutLearn can train the neural networks with comparable or even higher accuracy compared to traditional centralized/federated training mode with privacy preserved. AutLearn can also remarkably cut the computational and energy cost of neural network training on mobile devices.     

Convolutional neural network based early fire detection

The detection of manmade disasters particularly fire is valuable because it causes many damages in terms of human lives. Research on fire detection using wireless sensor network and video-based methods is a very hot research topic. However, the WSN based detection model need fire happens and a lot of smoke and fire for detection. Similarly, video-based models also have some drawbacks because conventional algorithms need feature vectors and high rule-based models for detection. In this paper, we proposed a fire detection method which is based on powerful machine learning and deep learning algorithms. We used both sensors data as well as images data for fire prevention. Our proposed model has three main deep neural networks i.e. a hybrid model which consists of Adaboost and many MLP neural networks, Adaboost-LBP model and finally convolutional neural network. We used Adaboost-MLP model to predict the fire. After the prediction, we proposed two neural networks i.e. Adaboost-LBP model and convolutional neural network for detection of fire using the videos and images taken from the cameras installed for the surveillance. Adaboost-LBP model is to generate the ROIs from the image where emergencies exist Our proposed model results are quite good, and the accuracy is almost 99%. The false alarming rate is very low and can be reduced more using further training.

     

高分辨卫星图像卷积神经网络分类模型

目的 卫星图像往往目标、背景复杂而且带有噪声,因此使用人工选取的特征进行卫星图像的分类就变得十分困难。提出一种新的使用卷积神经网络进行卫星图像分类的方案。使用卷积神经网络可以提取卫星图像的高层特征,进而提高卫星图像分类的识别率。方法 首先,提出一个包含六类图像的新的卫星图像数据集来解决卷积神经网络的有标签训练样本不足的问题。其次,使用了一种直接训练卷积神经网络模型和3种预训练卷积神经网络模型来进行卫星图像分类。直接训练模型直接在文章提出的数据集上进行训练,预训练模型先在ILSVRC（the ImageNet large scale visual recognition challenge）-2012数据集上进行预训练,然后在提出的卫星图像数据集上进行微调训练。完成微调的模型用于卫星图像分类。结果 提出的微调预训练卷积神经网络深层模型具有最高的分类正确率。在提出的数据集上,深层卷积神经网络模型达到了99.50%的识别率。在数据集UC Merced Land Use上,深层卷积神经网络模型达到了96.44%的识别率。结论 本文提出的数据集具有一般性和代表性,使用的深层卷积神经网络模型具有很强的特征提取能力和分类能力,且是一种端到端的分类模型,不需要堆叠其他模型或分类器。在高分辨卫星图像的分类上,本文模型和对比模型相比取得了更有说服力的结果。     

基于FR-ResNet的车辆型号精细识别研究

车辆型号精细识别的关键是提取有区分性的细节特征.以"特征重用"为核心,以有效提取车辆图像细节特征并进行高效利用为目的,提出了一种基于残差网络特征重用的深度卷积神经网络模型FR-ResNet(Improved ResNet focusing on feature reuse).该网络以ResNet残差结构为基础,分别采用...     

基于迭代训练和集成学习的图像分类方法

针对卷积神经网络提取特征信息不完整导致图像分类方法分类精度不高等问题,利用深度学习的方法搭建卷积神经网络模型框架,提出一种基于迭代训练和集成学习的图像分类方法。利用数据增强对图像数据集进行预处理操作,在提取图像特征时,采用一种迭代训练卷积神经网络的方式,得到充分有效的图像特征,在训练分类器时,采用机器学习中集成学习的思想。分别在特征提取后训练分类器,根据各分类器贡献的大小,赋予它们不同的权重值,取得比单个分类器更好的性能,提高图像分类的精度。该方法在Stanford Dogs、UEC FOOD-100和CIFAR-100数据集上的实验结果表明了其较好的分类性能。     

结合GAN的轻量级模糊车牌识别算法

目的 模糊车牌识别是车牌识别领域的难题,针对模糊车牌图像收集困难、车牌识别算法模型太大、不适用于移动或嵌入式设备等不足,本文提出了一种轻量级的模糊车牌识别方法,使用深度卷积生成对抗网络生成模糊车牌图像,用于解决现实场景中模糊车牌难以收集的问题,在提升算法识别准确性的同时提升了部署泛化能力。方法 该算法主要包含两部分,即基于优化卷积生成对抗网络的模糊车牌图像生成和基于深度可分离卷积网络与双向长短时记忆（long short-term memory,LSTM）的轻量级车牌识别。首先,使用Wasserstein距离优化卷积生成对抗网络的损失函数,提高生成车牌图像的多样性和稳定性;其次,在卷积循环神经网络的基础上,结合深度可分离卷积设计了一个轻量级的车牌识别模型,深度可分离卷积网络在减少识别算法计算量的同时,能对训练样本进行有效的特征学习,将特征图转换为特征序列后输入到双向LSTM网络中,进行序列学习与标注。结果 实验表明,增加生成对抗网络生成的车牌图像,能有效提高本文算法、传统车牌识别和基于深度学习的车牌识别方法的识别率,为进一步提高各类算法的识别率提供了一种可行方案。结合深度可分离卷积的轻量级车牌识别模型,识别率与基于标准循环卷积神经网络（convolutional recurrent neural network,CRNN）的车牌识别方法经本文生成图像提高后的识别率相当,但在模型的大小和识别速度上都优于标准的CRNN模型,本文算法的模型大小为45 MB,识别速度为12.5帧/s,标准CRNN模型大小是82 MB,识别速度只有7帧/s。结论 使用生成对抗网络生成图像,可有效解决模糊车牌图像样本不足的问题;结合深度可分离卷积的轻量级车牌识别模型,具有良好的识别准确性和较好的部署泛化能力。     

FedSharing：一种双区块链激励驱动的数据分享联邦学习框架

联邦学习（federated learning,FL）能够在不丢失数据所有权的同时依托隐私保护技术实现安全的分布式模型训练,但是其也具有中心化、缺乏公平激励等问题。区块链（blockchain）本质上来说是一种分布式数据库,具有去中心化、信任公证等特点,但是其也具有网络吞吐量小、资源浪费等关键问题。针对上述技术方法的问题与特点,提出了一种双区块链激励驱动的数据分享联邦学习框架,称为FedSharing。分别构建主链与侧链：主链使用交易封装联邦学习中交换的全局参数,同时结合链上智能合约和链下扩容技术建立梯度状态通道;侧链提出了一种新型的修正Shapley值工作量证明算法（PoFS）,修正传统Shapley值计算中成员平等性前提,将联邦学习中成员合作历史诚信度这一影响联盟利益的因素纳入考量。测试结果表明：梯度状态通道较智能合约去中心化方案每轮次时间平均降低4~5 s,PoFS共识下激励分配比例更符合公平实际。     

Times-series data augmentation and deep learning for construction equipment activity recognition

Automated, real-time, and reliable equipment activity recognition on construction sites can help to minimize idle time, improve operational efficiency, and reduce emissions. Previous efforts in activity recognition of construction equipment have explored different classification algorithms anm accelerometers and gyroscopes. These studies utilized pattern recognition approaches such as statistical models (e.g., hidden-Markov models); shallow neural networks (e.g., Artificial Neural Networks); and distance algorithms (e.g., K-nearest neighbor) to classify the time-series data collected from sensors mounted on the equipment. Such methods necessitate the segmentation of continuous operational data with fixed or dynamic windows to extract statistical features. This heuristic and manual feature extraction process is limited by human knowledge and can only extract human-specified shallow features. However, recent developments in deep neural networks, specifically recurrent neural network (RNN), presents new opportunities to classify sequential time-series data with recurrent lateral connections. RNN can automatically learn high-level representative features through the network instead of being manually designed, making it more suitable for complex activity recognition. However, the application of RNN requires a large training dataset which poses a practical challenge to obtain from real construction sites. Thus, this study presents a data-augmentation framework for generating synthetic time-series training data for an RNN-based deep learning network to accurately and reliably recognize equipment activities. The proposed methodology is validated by generating synthetic data from sample datasets, that were collected from two earthmoving operations in the real world. The synthetic data along with the collected data were used to train a long short-term memory (LSTM)-based RNN. The trained model was evaluated by comparing its performance with traditionally used classification algorithms for construction equipment activity recognition. The deep learning framework presented in this study outperformed the traditionally used machine learning classification algorithms for activity recognition regarding model accuracy and generalization.     

基于凸优化的无人驾驶汽车转向角安全性验证

无人驾驶汽车系统过大的输入-输出空间(即输入和输出的所有可能组合)使得为其提供形式化保证变成一项具有挑战性的任务.在本文中,我们提出了一种自动验证技术,通过结合凸优化和深度学习验证工具DLV来保障无人驾驶汽车的转向角安全.DLV是一个用于自动验证图像分类神经网络安全性的框架.我们运用故障安全轨迹规划中的凸优化技术解决预测转向角的判断问题,然后拓展DLV来实现无人驾驶汽车转向角安全性的验证.我们在NVIDIA的端到端无人驾驶架构上说明所提出方法的优势,这个架构是许多现代无人驾驶汽车的关键组成部分.我们的实验结果表明,对于给定区域和操作集,如果存在对抗性错误分类(即不正确的转向决策),我们的技术可以成功地找到.因此,我们可以实现安全验证(如果在所有DNN层都没有发现错误分类,在这种情况下网络关于转向决策可以说是稳定或可靠的)或证伪(在这种情况下,这些对抗性反例可以用于后续微调网络).     

基于边界值不变量的对抗样本检测方法

目前,深度学习成为计算机领域研究与应用最广泛的技术之一,在图像识别、语音、自动驾驶、文本翻译等方面都取得良好的应用成果。但人们逐渐发现深度神经网络容易受到微小扰动图片的影响,导致分类出现错误,这类攻击手段被称为对抗样本。对抗样本的出现可能会给安全敏感的应用领域带来灾难性的后果。现有的防御手段大多需要对抗样本本身作为训练集,这种对抗样本相关的防御手段是无法应对未知对抗样本攻击的。借鉴传统软件安全中的边界检查思想,提出了一种基于边界值不变量的对抗样本检测防御方法,该方法通过拟合分布来寻找深度神经网络中的不变量,且训练集的选取与对抗样本无关。实验结果表明,在 LeNet、vgg19 模型和 Mnist、Cifar10 数据集上,与其他对抗检测方法相比,提出的方法可有效检测目前的常见对抗样本攻击,并且具有低误报率。     

基于小样本学习的SAR图像识别

深度学习已成为图像识别领域的一个研究热点。与传统图像识别方法不同,深度学习从大量数据中自动学习特征,并且具有强大的自学习能力和高效的特征表达能力。但在小样本条件下,传统的深度学习方法如卷积神经网络难以学习到有效的特征,造成图像识别的准确率较低。因此,提出一种新的小样本条件下的图像识别算法用于解决SAR图像的分类识别。该算法以卷积神经网络为基础,结合自编码器,形成深度卷积自编码网络结构。首先对图像进行预处理,使用2D Gabor滤波增强图像,在此基础上对模型进行训练,最后构建图像分类模型。该算法设计的网络结构能自动学习并提取小样本图像中的有效特征,进而提高识别准确率。在MSTAR数据集的10类目标分类中,选择训练集数据中10%的样本作为新的训练数据,其余数据为验证数据,并且,测试数据在卷积神经网络中的识别准确率为76.38%,而在提出的卷积自编码结构中的识别准确率达到了88.09%。实验结果表明,提出的算法在小样本图像识别中比卷积神经网络模型更加有效。     

Research on Transfer Learning Methods for Classification of Typhoon Cloud Image

Aiming at the complexity of traditional methods for feature extraction about satellite cloud images, and the difficulty of developing deep convolutional neural network from scratch, a parameter-based transfer learning method for classifying typhoon intensity is proposed. Take typhoon satellite cloud images published by Japan Meteorological Agency, which includes 10 000 scenes among nearly 40 years to construct training and test typhoon datasets. Three deep convolutional neural networks, VGG16, InceptionV3 and ResNet50 are trained as source models on the large-scale ImageNet datasets. Considering the discrepancy between low-level features and high-level semantic features of typhoon cloud images, adapt the optimal number of transferable layers in neural networks and freeze weights of low-level network. Meanwhile, fine-tune surplus weights on typhoon dataset adaptively. Finally, a transferred prediction model which is suitable for small sample typhoon datasets, called T-typCNNs is proposed. Experimental results show that the T-typCNNs can achieve training accuracy of 95.081% and testing accuracy of 91.134%, 18.571% higher than using shallow convolutional neural network, 9.819% higher than training with source models from scratch.     

基于对抗学习和知识蒸馏的神经网络压缩算法

针对基于深度学习的人脸识别模型难以在嵌入式设备进行部署和实时性能差的问题,深入研究了现有的模型压缩和加速算法,提出了一种基于知识蒸馏和对抗学习的神经网络压缩算法。算法框架由三部分组成,预训练的大规模教师网络、轻量级的学生网络和辅助对抗学习的判别器。改进传统的知识蒸馏损失,增加指示函数,使学生网络只学习教师网络正确识别的分类概率;鉴于中间层特征图具有丰富的高维特征,引入对抗学习策略中的判别器,鉴别学生网络与教师网络在特征图层面的差异;为了进一步提高学生网络的泛化能力,使其能够应用于不同的机器视觉任务,在训练的后半部分教师网络和学生网络相互学习,交替更新,使学生网络能够探索自己的最优解空间。分别在CASIA WEBFACE和CelebA两个数据集上进行验证,实验结果表明知识蒸馏得到的小尺寸学生网络相较全监督训练的教师网络,识别准确率仅下降了1.5%左右。同时将本研究所提方法与面向特征图知识蒸馏算法和基于对抗学习训练的模型压缩算法进行对比,所提方法具有较高的人脸识别准确率。     

基于分散神经网络的多指灵巧手位置控制

针对多指灵巧手钢缆传动系统的非线性，提出一种基于分散神经网络的位置控制方法．通过 对复杂的钢缆传动系统施加不同的输入可以得到特定的相对简单的输入输出数据，利用这种 特定的输入输出数据学习传动系统的非线性关系得到多个分散的神经网络，再根据传动系统 的结构特性用分散的神经网络求取钢缆传动系统的逆模型，用于直接逆控制，从而达到补偿 非线性误差的目的．同时应用在线神经网络的适时补偿使系统长时间保持良好的运行状态． 实验证明这种方法可大大提高位置跟踪精度，取得比较满意的结果．     

Quality-distinguishing and patch-comparing no-reference image quality assessment

No-reference image quality assessment (NR-IQA) based on deep learning attracts a great research attention recently. However, its performance in terms of accuracy and efficiency is still under exploring. To address these issues, in this paper, we propose a quality-distinguishing and patch-comparing NR-IQA approach based on convolutional neural network (QDPC-CNN). We improve the prediction accuracy by two proposed mechanisms: quality-distinguishing adaption and patch-comparing regression. The former trains multiple models from different subsets of a dataset and adaptively selects one for predicting quality score of a test image according to its quality level, and the latter generates patch pairs for regression under different combination strategies to make better use of reference images in network training and enlarge training data at the same time. We further improve the efficiency of network training by a new patch sampling way based on the visual importance of each patch. We conduct extensive experiments on several public databases and compare our proposed QDPC-CNN with existing state-of-the-art methods. The experimental results demonstrate that our proposed method outperforms the others both in terms of accuracy and efficiency.

     

基于迁移学习及气象卫星云图的台风等级分类研究

针对传统卫星云图特征提取方法复杂且深度卷积神经网络（Deep Convolutional Neural Network, DCNN）模型开发困难的问题,提出一种基于参数迁移的台风等级分类方法。利用日本气象厅发布的近40 a 10 000多景台风云图数据,构建了适应于迁移学习的台风云图训练集和测试集。在大规模ImageNet源数据集上训练出3种源模型VGG16,InceptionV3和ResNet50,依据台风云图低层特征与高层语义特征的差异,适配网络最佳迁移层数并冻结低层权重,高层权重采用自适应微调策略,构建出了适用于台风小样本数据集的迁移预报模型T-typCNNs。实验结果表明:T-typCNNs模型在自建台风数据集上的训练精度为95.081%,验证精度可达91.134%,比利用浅层卷积神经网络训练出的精度高18.571%,相比于直接用源模型训练最多提高9.819%。