全卷积多并联残差神经网络

随着人工智能的火热发展,深度学习已经在很多领域占有了一席之地.作为深度学习中一个典型网络--残差神经网络模型自提出之日起就成为了众多研究者的关注点.然而,残差神经网络还有很大的改进空间.为了更好地解决反向传播中梯度减小的问题,本文提出了一种改进的残差神经网络,称为全卷积多并联残差神经网络.在该网络中,每一层的特征信息不仅传输到下一层还输出到最后的平均池化层.为了测试该网络的性能,分别在三个数据集(MNIST,CIFAR-10和CIFAR-100)上对比图像分类的结果.实验结果表明,改进后的全卷积多并联残差神经网络与残差网络相比具有更高的分类准确率和更好的泛化能力.     

改进的轻量型网络在图像识别上的应用

针对卷积神经网络在图像识别任务上模型复杂度大、参数量多,首先提出了一种轻量化的SepNet网络结构,该结构在分类器模块上采用克罗内克积替换了传统的全连接层。为进一步优化网络结构,在特征提取模块均衡网络深度、宽度,设计了一个利用深度可分离卷积和残差网络的可分离残差模块,最终形成了一个能实现端到端训练的轻量化网络架构,称为sep_res18_s3。实验分别在MNIST、CIFAR-10、CIFAR-100数据集上验证SepNet的有效性,设计的SepNet网络结构相比VGG10,参数数量和运算量在不损失其精度下均降低了94.15%。同时,相比设计的类残差网络cov_res18_s3,sep_res18_s3仍能降低58.33%的参数量和81.82%的FLOPs。实验结果表明,采用克罗内克积替换全连接层可以在保证训练结果准确度的同时显著降低参数数量和计算成本,并在一定程度上防止过拟合,在此基础上结合深度可分离卷积和类残差结构,证明了sep_res18_s3的有效性。     

基于改进残差网络的人脸识别算法

针对传统的人脸识别技术对于人脸特征提取的能力有限的问题,提出了一种改进的残差神经网络人脸识别算法.通过对原有残差神经网络模型结构的调整,适当增加卷积输出,减少残差单元的方法来提高网络性能,从而提高残差神经网络提取人脸特征的能力.实验结果表明:提出的算法在自建数据集OurFace和CASIA-WebFace数据集上取得了优于现有残差网络的人脸识别性能.     

基于稀疏正则化的卷积神经网络模型剪枝方法

现有卷积神经网络模型剪枝方法仅依靠自身参数信息难以准确评估参数重要性,容易造成参数误剪且影响网络模型整体性能。提出一种改进的卷积神经网络模型剪枝方法,通过对卷积神经网络模型进行稀疏正则化训练,得到参数较稀疏的深度卷积神经网络模型,并结合卷积层和BN层的稀疏性进行结构化剪枝去除冗余滤波器。在CIFAR-10、CIFAR-100和SVHN数据集上的实验结果表明,该方法能有效压缩网络模型规模并降低计算复杂度,尤其在SVHN数据集上,压缩后的VGG-16网络模型在参数量和浮点运算量分别减少97.3%和91.2%的情况下,图像分类准确率仅损失了0.57个百分点。     

一种并行不对称空洞卷积模块

构建卷积神经网络要耗费大量的人力资源,且训练过程中需要消耗大量的算力资源.利用空洞卷积代替卷积神经网络中的池化操作,能有效增加感受野,降低运算复杂度,但是空洞卷积会带来空间层次和信息连续性的丢失.本文提出了一种并行不对称空洞卷积模块,该模块能够补全空洞卷积所丢失的信息,可以嵌入到现有的卷积神经网络中,代替3×3卷积进行网络训练,从而加速网络的收敛,提高网络的性能.实验结果表明,利用本文所提出的并行不对称空洞卷积模块,可以显著提高不同网络在CIFAR-10等数据集上的分类效果.     

基于残差的优化卷积神经网络服装分类算法

针对目前服装分类算法在解决多类别服装分类问题时分类精度一般的问题,提出了一种基于残差的优化卷积神经网络服装分类算法,在网络中使用了如下三种优化方法:(1)调整批量归一化层、激活函数层与卷积层在网络中的排列顺序;(2)"池化层+卷积层"的并行池化结构;(3)使用全局均值池化层替换全连接层。经过由香港中文大学多媒体实验室提供的多类别大型服装数据集(DeepFashion)和标准数据集CIFAR-10上的实验表明,所提出的网络模型在处理图片的速度和分类精度方面都优于VGGNet和AlexNet,且得到了目前为止已知的在DeepFashion数据集上最好的分类准确率。该网络也可以更好地应用于目标检测和图像分割领域。     

自适应融合残差网在图像分类中应用研究

针对卷积神经网络存在随着网络深度增加导致优化困难,识别正确率降低、泛化性能差等问题,在Res Net(残差网络)基础上,提出了一种基于softmax全连接自适应门控网络融合模型.该方法在隐层网络深度达到一定层数后,设置多种卷积核尺寸作为独立网络输出,通过softmax全连接门控网络输出各模型选择概率,融合多种卷积尺寸残差网输出作为模型最终输出.实验表明,本文提出的融合残差网络模型更适合于多类别、精细化数据集,与单网络模型相比,在训练集上具有更好的收敛性,在测试集上具有更好的泛化性能.     

一种改进残差网络的服装图像识别模型

传统的服装多类别分类方法主要是人工提取图像的颜色、纹理、边缘等特征,这些人工选取特征方法过程繁琐且分类精度较低。深度残差网络可通过增加神经网络的深度获得较高的识别精度被广泛地应用于各个领域。为提高服装图像识别精度问题,提出一种改进深度残差网络模型：改进残差块中卷积层、调整批量归一化层与激活函数层中的排列顺序;引入注意力机制;调整网络卷积核结构。该网络结构在标准数据集Fashion-MNIST和香港中文大学多媒体实验室提供的多类别大型服装数据集（DeepFashion）上进行测试,实验结果表明,所提出的网络模型在服装图像识别分类精度上优于传统的深度残差网络。     

基于无效卷积核权值回退的神经网络优化方法

在神经网络模型训练过程中,存在部分卷积核退化为无效卷积核,在神经网络推理过程失去作用的问题。针对该问题,提出了一种仅使用单个模型就能在训练过程中激活无效卷积核,提高模型性能的方法。首先将初始模型训练至收敛时刻;然后通过L1正则和卷积核相关性两种方式衡量卷积核的有效性;最后将无效卷积核的权值回退到模型训练的初期阶段并对模型进行重训练。在CIFAR-10、CIFAR-100等图像分类的数据集上的实验结果表明,无论是在残差网络还是在轻量级网络上,提出的方法都能有效地恢复无效卷积核,提高神经网络模型精度。相比之前的方法,该方法在低代价下达到了最佳效果,在图像分类任务上平均提高了0.93%的准确率。     

双分支多注意力机制的锐度感知分类网络

基于卷积神经网络的图像分类方法的关键是提取有区分性的重点特征.为了提高重点特征的关注度，增强网络泛化能力，文中提出双分支多注意力机制的锐度感知分类网络(Double-Branch Multi-attention Mechanism Based Sharpness-Aware Classification Network, DAMSNet).该网络以ResNet-34残差网络为基础，首先，修改ResNet-34残差网络输入层卷积核尺寸，删除最大池化层，减小原始图像特征的损失.再者，提出双分支多注意力机制模块，嵌入残差分支中，从全局特征和局部特征上提取图像在通道域和空间域的上下文信息.然后，引入锐度感知最小化算法，结合随机梯度下降优化器，同时最小化损失值和损失锐度，寻找具有一致低损失的邻域参数，提高网络泛化能力.在CIFAR-10、CIFAR-100、SVHN数据集上的实验表明，文中网络不仅具有较高的分类精度，而且有效提升泛化能力.     

SRNET: A Shallow Skip Connection Based Convolutional Neural Network Design for Resolving Singularities

Convolutional neural networks (CNNs) have shown tremendous progress and performance in recent years. Since emergence, CNNs have exhibited excellent performance in most of classification and segmentation tasks. Currently, the CNN family includes various architectures that dominate major vision-based recognition tasks. However, building a neural network (NN) by simply stacking convolution blocks inevitably limits its optimization ability and introduces overfitting and vanishing gradient problems. One of the key reasons for the aforementioned issues is network singularities, which have lately caused degenerating manifolds in the loss landscape. This situation leads to a slow learning process and lower performance. In this scenario, the skip connections turned out to be an essential unit of the CNN design to mitigate network singularities. The proposed idea of this research is to introduce skip connections in NN architecture to augment the information flow, mitigate singularities and improve performance. This research experimented with different levels of skip connections and proposed the placement strategy of these links for any CNN. To prove the proposed hypothesis, we designed an experimental CNN architecture, named as Shallow Wide ResNet or SRNet, as it uses wide residual network as a base network design. We have performed numerous experiments to assess the validity of the proposed idea. CIFAR-10 and CIFAR-100, two well-known datasets are used for training and testing CNNs. The final empirical results have shown a great many of promising outcomes in terms of performance, efficiency and reduction in network singularities issues.

     

基于改进深度残差网络的低功耗表情识别

为了提高表情识别率并降低表情识别的功耗,提出一种基于改进深度残差网络的表情识别方法。残差学习在解决深度卷积神经网络退化问题、使网络层次大幅加深的同时,进一步增加了网络的功耗。为此,引入具有生物真实性的激活函数来代替已有的整流线性单元(Rectified Linear Units,ReLU)函数, 并将其作为卷积层激活函数对深度残差网络进行改进。该方法不仅提高了残差网络的精度,而且训练出的网络权重可直接作为与该深度残差网络具有相同结构的深度脉冲神经网络的权重。将该深度脉冲神经网络部署在类脑硬件上时,其能够以较高的识别率和较低的能耗进行表情识别。     

A novel group squeeze excitation sparsely connected convolutional networks for SAR target classification

Automatic Target Recognition (ATR) based on Synthetic Aperture Radar (SAR) images plays a key role in military applications. However, there are difficulties with this traditional recognition method. Principally, it is a challenge to design robust features and classifiers for different SAR images. Although Convolutional Neural Networks (CNNs) are very successful in many image classification tasks, building a deep network with limited labeled data remains a problem. The topologies of CNNs like the fully connected structure will lead to redundant parameters and the negligence of channel-wise information flow. A novel CNNs approach, called Group Squeeze Excitation Sparsely Connected Convolutional Networks (GSESCNNs), is therefore proposed as a solution. The group squeeze excitation performs dynamic channel-wise feature recalibration with less parameters than squeeze excitation. Sparsely connected convolutional networks are a more efficient way to operate the concatenation of feature maps from different layers. Experimental results on Moving and Stationary Target Acquisition and Recognition (MSTAR) SAR images, demonstrate that this approach achieves, at 99.79%, the best prediction accuracy, outperforming the most common skip connection models, such as Residual Networks and Densely Connected Convolutional Networks, as well as other methods reported in the MSTAR dataset.     

基于改进Fast-SCNN的塑瓶气泡缺陷实时分割算法

在医用塑瓶的瓶身气泡检测时，瓶身气泡位置的任意性、气泡大小的不确定性以及气泡特征与瓶身特征之间的相似性增加了气泡缺陷的检测难度。针对上述气泡缺陷检测难点问题，提出了一种基于改进快速分割卷积神经网络(Fast-SCNN)的实时分割算法。该分割算法的基础框架为Fast-SCNN，而为弥补原有网络分割尺寸的鲁棒性不足，借鉴了SENet的通道间信息的利用与多级跳跃连接的思想，具体为网络进一步下采样提取深层特征，在解码阶段将上采样操作融合SELayer模块，同时增加两次与网络浅层的跳跃连接。设计四组对比实验，在气泡数据集上以平均交并比（MIoU）与算法单张分割时间作为评价指标。实验结果表明，改进Fast-SCNN的综合性能最好，其MIoU为97.08%，其预处理后的医用塑瓶的平均检测时间为24.4 ms，其边界分割准确率较Fast-SCNN提升了2.3%，增强了对微小气泡的分割能力，而且该网络的MIoU相较现有的U-Net提升了0.27%，时间上降低了7.5 ms，综合检测性能远超过全卷积神经网络(FCN-8s)。该算法能够有效地对较小的、边缘不清晰的气泡进行分割，满足对气泡缺陷实时分割检测的工程要求。     

基于递归卷积神经网络的移动机器人定位算法

移动机器人定位已成为机器人研究的重要任务。提出基于递归卷积神经网络的移动机器人定位（Recurrent Convolutional Neural Networks-Based Mobile Robot Localization,RCNN-MRL）算法。递归卷积神经网络（Recurrent Convolutional Neural Networks,RCNN）结合卷积神经网络（Convolutional Neural Networks,CNN）和递归神经网络（Recurrent Neural Networks,RNN）的特性,并依据机器人上嵌入的照相机拍摄的第一人称视角图像,RCNN-MRL算法利用RCNN实现自主定位。具体而言,先通过RCNN有效地处理多个连续图像,再利用RCNN作为回归模型,进而估计机器人位置。同时,设计双轮机器人移动,获取多个时间序列图像信息。最后,依据双轮机器人随机移动建立仿真环境,分析机器人定位性能。实验数据表明,提出的RCNN模型能够实现自主定位。     

Deep topology network: A framework based on feedback adjustment learning rate for image classification

Convolutional Neural Network (CNN) has demonstrated its superior ability to achieve amazing accuracy in computer vision field. However, due to the limitation of network depth and computational complexity, it is still difficult to obtain the best classification results for the specific image classification tasks. In order to improve classification performance without increasing network depth, a new Deep Topology Network (DTN) framework is proposed. The key idea of DTN is based on the iteration of multiple learning rate feedback. The framework consists of multiple sub-networks and each sub-network has its own learning rate. After the determined iteration period, these learning rates can be adjusted according to the feedback of training accuracy, in the feature learning process, the optimal learning rate is updated iteratively to optimize the loss function. In practice, the proposed DTN framework is applied to several state-of-the-art deep networks, and its performance is tested by extensive experiments and comprehensive evaluations of CIFAR-10 and MNIST benchmarks. Experimental results show that most deep networks can benefit from the DTN framework with an accuracy of 99.5% on MINIST dataset, which is 5.9% higher than that on the CIFAR-10 benchmark.     

基于卷积词袋网络的视觉识别

近年来,卷积神经网络（CNN）凭借其强大的特征学习能力在视觉识别领域取得重要进展。针对CNN全连接层对图像平移、旋转、缩放等变换比较敏感的问题,提出了一种混合模型--卷积词袋网络（BoCW-Net）。它将BoW模型嵌入CNN结构中并代替全连接层,通过端到端的方式学习特征、字典和分类器。为实现BoCW-Net整个网络的有监督学习,提出基于方向相似度的BoCW编码。同时,为充分利用中层特征和高层特征的鉴别性,将中层辅助分类器与高层分类器集成,形成主-辅集成分类器。实验结果表明：相比全连接层,BoCW表示对各种变换具有更强的不变性;主-辅集成分类器能有效融合中层、高层特征,提高BoCW-Net的识别性能;相比新近发展的CNN模型,BoCW-Net在CIFAR-10、CIFAR-100和MNIST数据库上均取得了改进的识别性能,最终分别获得4.88%、22.48%和0.21%的测试错误率。     

Image super-resolution based on two-level residual learning CNN

In recent years, CNN has been used for single image super-resolution (SR) with its success of in the field of computer vision. However, in the recovery process, there are always some high-frequency components that cant be recovered from low-resolution images to high-resolution ones by using existing CNN-based methods. In this paper, we propose an image super-resolution method based on CNN, which uses a two-level residual learning network to learn residual components, i.e., high-frequency components. We use the Super-Resolution Convolutional Neural Network (SRCNN) as the network structure in each level so that our proposed method can achieve the high-resolution images with high-frequency components that cant be obtained by the existing methods. In addition, we analyze the proposed method with considering three kinds of residual learning networks, which are different in the structure and superimposed layers of the residual learning network. In the experiments, we investigate the performance of the proposed method with various residual learning networks and the effect of image super-resolution to image captioning task.

     

Convolutional long short-term memory model for recognizing construction workers’ postures from wearable inertial measurement units

This paper proposes using Deep Neural Networks (DNN) models for recognizing construction workers’ postures from motion data captured by wearable Inertial Measurement Units (IMUs) sensors. The recognized awkward postures can be linked to known risks of Musculoskeletal Disorders among workers. Applying conventional Machine Learning (ML)-based models has shown promising results in recognizing workers’ postures. ML models are limited – they reply on heuristic feature engineering when constructing discriminative features for characterizing postures. This makes further improving the model performance regarding recognition accuracy challenging. In this paper, the authors investigate the feasibility of addressing this problem using a DNN model that, through integrating Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) layers, automates feature engineering and sequential pattern detection. The model’s recognition performance was evaluated using datasets collected from four workers on construction sites. The DNN model integrating one convolutional and two LSTM layers resulted in the best performance (measured by F1 Score). The proposed model outperformed baseline CNN and LSTM models suggesting that it leveraged the advantages of the two baseline models for effective feature learning. It improved benchmark ML models’ recognition performance by an average of 11% under personalized modelling. The recognition performance was also improved by 3% when the proposed model was applied to 8 types of postures across three subjects. These results support that the proposed DNN model has a high potential in addressing challenges for improving the recognition performance that was observed when using ML models.