Accurate and Robust Eye Center Localization via Fully Convolutional Networks

Eye center localization is one of the most crucial and basic requirements for some human-computer interaction applications such as eye gaze estimation and eye tracking. There is a large body of works on this topic in recent years, but the accuracy still needs to be improved due to challenges in appearance such as the high variability of shapes, lighting conditions, viewing angles and possible occlusions. To address these problems and limitations, we propose a novel approach in this paper for the eye center localization with a fully convolutional network(FCN),which is an end-to-end and pixels-to-pixels network and can locate the eye center accurately. The key idea is to apply the FCN from the object semantic segmentation task to the eye center localization task since the problem of eye center localization can be regarded as a special semantic segmentation problem. We adapt contemporary FCN into a shallow structure with a large kernel convolutional block and transfer their performance from semantic segmentation to the eye center localization task by fine-tuning.Extensive experiments show that the proposed method outperforms the state-of-the-art methods in both accuracy and reliability of eye center localization. The proposed method has achieved a large performance improvement on the most challenging database and it thus provides a promising solution to some challenging applications.

Segmentation of retinal fluid based on deep learning:application of three-dimensional fully convolutional neural networks in optical coherence tomography images

AIM: To explore a segmentation algorithm based on deep learning to achieve accurate diagnosis and treatment of patients with retinal fluid.METHODS: A two-dimensional(2D) fully convolutional network for retinal segmentation was employed. In order to solve the category imbalance in retinal optical coherence tomography(OCT) images, the network parameters and loss function based on the 2D fully convolutional network were modified. For this network, the correlations of corresponding positions among adjacent images in space are ignored. Thus, we proposed a three-dimensional(3D) fully convolutional network for segmentation in the retinal OCT images.RESULTS: The algorithm was evaluated according to segmentation accuracy, Kappa coefficient, and F1 score. For the 3D fully convolutional network proposed in this paper, the overall segmentation accuracy rate is 99.56%, Kappa coefficient is 98.47%, and F1 score of retinal fluid is 95.50%. CONCLUSION: The OCT image segmentation algorithm based on deep learning is primarily founded on the 2D convolutional network. The 3D network architecture proposed in this paper reduces the influence of category imbalance, realizes end-to-end segmentation of volume images, and achieves optimal segmentation results. The segmentation maps are practically the same as the manual annotations of doctors, and can provide doctors with more accurate diagnostic data.

Prediction of Uncertainty Estimation and Confidence Calibration Using Fully Convolutional Neural Network

Convolution neural networks(CNNs)have proven to be effective clinical imagingmethods.This study highlighted some of the key issues within these systems.It is difficult to train these systems in a limited clinical image databases,and many publications present strategies including such learning algorithm.Furthermore,these patterns are known formaking a highly reliable prognosis.In addition,normalization of volume and losses of dice have been used effectively to accelerate and stabilize the training.Furthermore,these systems are improperly regulated,resulting in more confident ratings for correct and incorrect classification,which are inaccurate and difficult to understand.This study examines the risk assessment of Fully Convolutional Neural Networks(FCNNs)for clinical image segmentation.Essential contributions have been made to this planned work:1)dice loss and cross-entropy loss are compared on the basis of segment quality and uncertain assessment of FCNNs;2)proposal for a group model for assurance measurement of full convolutional neural networks trained with dice loss and group normalization;And 3)the ability of the measured FCNs to evaluate the segment quality of the structures and to identify test examples outside the distribution.To evaluate the study’s contributions,it conducted a series of tests in three clinical image division applications such as heart,brain and prostate.The findings of the study provide significant insights into the predictive ambiguity assessment and a practical strategies for outside-distribution identification and reliable measurement in the clinical image segmentation.The approaches presented in this research significantly enhance the reliability and accuracy rating of CNNbased clinical imaging methods.

Fully Convolutional Networks for Street Furniture Identification in Panorama Images

Panoramic images are widely used in many scenes,especially in virtual reality and street view capture.However,they are new for street furniture identification which is usually based on mobile laser scanning point cloud data or conventional 2D images.This study proposes to perform semantic segmentation on panoramic images and transformed images to separate light poles and traffic signs from background implemented by pre-trained Fully Convolutional Networks(FCN).FCN is the most important model for deep learning applied on semantic segmentation for its end to end training process and pixel-wise prediction.In this study,we use FCN-8s model that pre-trained on cityscape dataset and finetune it by our own data.Then replace cross entropy loss function with focal loss function in the FCN model and train it again to produce the predictions.The results show that in all results from pre-trained model,fine-tuning,and FCN model with focal loss,the light poles and traffic signs are detected well and the transformed images have better performance than panoramic images in the prediction according to the Recall and IoU evaluation.

Automated Delineation of Smallholder Farm Fields Using Fully Convolutional Networks and Generative Adversarial Networks

Accurate boundaries of smallholder farm fields are important and indispensable geo-information that benefits farmers,managers,and policymakers in terms of better managing and utilizing their agricultural resources.Due to their small size,irregular shape,and the use of mixed-cropping techniques,the farm fields of smallholder can be difficult to delineate automatically.In recent years,numerous studies on field contour extraction using a deep Convolutional Neural Network(CNN)have been proposed.However,there is a relative shortage of labeled data for filed boundaries,thus affecting the training effect of CNN.Traditional methods mostly use image flipping,and random rotation for data augmentation.In this paper,we propose to apply Generative Adversarial Network(GAN)for the data augmentation of farm fields label to increase the diversity of samples.Specifically,we propose an automated method featured by Fully Convolutional Neural networks(FCN)in combination with GAN to improve the delineation accuracy of smallholder farms from Very High Resolution(VHR)images.We first investigate four State-Of-The-Art(SOTA)FCN architectures,i.e.,U-Net,PSPNet,SegNet and OCRNet,to find the optimal architecture in the contour detection task of smallholder farm fields.Second,we apply the identified optimal FCN architecture in combination with Contour GAN and pixel2pixel GAN to improve the accuracy of contour detection.We test our method on the study area in the Sudano-Sahelian savanna region of northern Nigeria.The best combination achieved F1 scores of 0.686 on Test Set 1(TS1),0.684 on Test Set 2(TS2),and 0.691 on Test Set 3(TS3).Results indicate that our architecture adapts to a variety of advanced networks and proves its effectiveness in this task.The conceptual,theoretical,and experimental knowledge from this study is expected to seed many GAN-based farm delineation methods in the future.

Intelligent Detection Model Based on a Fully Convolutional Neural Network for Pavement Cracks

The crack is a common pavement failure problem.A lack of periodic maintenance will result in extending the cracks and damage the pavement,which will affect the normal use of the road.Therefore,it is significant to establish an efficient intelligent identification model for pavement cracks.The neural network is a method of simulating animal nervous systems using gradient descent to predict results by learning a weight matrix.It has been widely used in geotechnical engineering,computer vision,medicine,and other fields.However,there are three major problems in the application of neural networks to crack identification.There are too few layers,extracted crack features are not complete,and the method lacks the efficiency to calculate the whole picture.In this study,a fully convolutional neural network based on ResNet-101 is used to establish an intelligent identification model of pavement crack regions.This method,using a convolutional layer instead of a fully connected layer,realizes full convolution and accelerates calculation.The region proposals come from the feature map at the end of the base network,which avoids multiple computations of the same picture.Online hard example mining and data-augmentation techniques are adopted to improve the model’s recognition accuracy.We trained and tested Concrete Crack Images for Classification(CCIC),which is a public dataset collected using smartphones,and the Crack Image Database(CIDB),which was automatically collected using vehicle-mounted charge-coupled device cameras,with identification accuracy reaching 91.4%and 86.4%,respectively.The proposed model has a higher recognition accuracy and recall rate than Faster RCNN and different depth models,and can extract more complete and accurate crack features in CIDB.We also analyzed translation processing,fuzzy,scaling,and distorted images.The proposed model shows a strong robustness and stability,and can automatically identify image cracks of different forms.It has broad application prospects in practical engineering problems.

A Fully Convolutional Neural Network-based Regression Approach for Effective Chemical Composition Analysis Using Near-infrared Spectroscopy in Cloud

As one chemical composition,nicotine content has an important influence on the quality of tobacco leaves.Rapid and nondestructive quantitative analysis of nicotine is an important task in the tobacco industry.Near-infrared(NIR)spectroscopy as an effective chemical composition analysis technique has been widely used.In this paper,we propose a one-dimensional fully convolutional network(1D-FCN)model to quantitatively analyze the nicotine composition of tobacco leaves using NIR spectroscopy data in a cloud environment.This 1D-FCN model uses one-dimensional convolution layers to directly extract the complex features from sequential spectroscopy data.It consists of five convolutional layers and two full connection layers with the max-pooling layer replaced by a convolutional layer to avoid information loss.Cloud computing techniques are used to solve the increasing requests of large-size data analysis and implement data sharing and accessing.Experimental results show that the proposed 1D-FCN model can effectively extract the complex characteristics inside the spectrum and more accurately predict the nicotine volumes in tobacco leaves than other approaches.This research provides a deep learning foundation for quantitative analysis of NIR spectral data in the tobacco industry.

Reconstructing the 3D digital core with a fully convolutional neural network

In this paper, the complete process of constructing 3D digital core by fullconvolutional neural network is described carefully. A large number of sandstone computedtomography (CT) images are used as training input for a fully convolutional neural networkmodel. This model is used to reconstruct the three-dimensional (3D) digital core of Bereasandstone based on a small number of CT images. The Hamming distance together with theMinkowski functions for porosity, average volume specifi c surface area, average curvature,and connectivity of both the real core and the digital reconstruction are used to evaluate theaccuracy of the proposed method. The results show that the reconstruction achieved relativeerrors of 6.26%, 1.40%, 6.06%, and 4.91% for the four Minkowski functions and a Hammingdistance of 0.04479. This demonstrates that the proposed method can not only reconstructthe physical properties of real sandstone but can also restore the real characteristics of poredistribution in sandstone, is the ability to which is a new way to characterize the internalmicrostructure of rocks.

基于全卷积网络的复杂背景红外弱小目标检测研究

针对复杂背景红外弱小目标检测过程中存在的检测误差率高,检测时间过长等问题,提出基于全卷积网络的复杂背景红外弱小目标检测方法。分析复杂背景红外弱小目标检测的研究进展,找出不同方法的缺陷,采集红外图像,提取目标检测特征,并采用全卷积网络设计弱小目标检测的分类器,实现复杂背景红外弱小目标检测。实验结果表明,该方法的复杂背景红外弱小目标检测精度超过97%,具有较高的实际应用价值。

基于LSTM-SAFCN模型的生物质锅炉NO_(x)排放浓度预测

针对生物质锅炉燃烧过程的动态特性,提出一种改进的长短期记忆-自注意力机制全卷积神经网络(LSTM-SAFCN)模型用于预测NO_(x)排放浓度。首先利用完全自适应噪声集合经验模态分解法(CEEMDAN)对数据进行预处理,消除数据噪声对NO_(x)排放浓度预测的影响;其次融合自注意力机制与长短时记忆-全卷积神经网络(LSTM-FCN)进行特征提取与预测建模,该拓展方法能够同时兼顾时间序列数据的局部细节与长期趋势特征;最后,利用生物质热电联产系统的实际运行数据验证了所提算法的有效性。

新工科背景下基于神经网络的隧道工程缺陷检测研究

隧道是综合交通运输体系的重要组成部分,是人民便捷生活的基本保障。为响应新工科建设号召,应对新一轮科技革命和产业变革,提升科学化建设要求,基于全卷积神经网络对隧道工程缺陷检测中表面裂缝检测方法进行了探究。通过已有照明拍摄平台采集到的隧道表面图片,采用全卷积神经网络模型对图像中裂缝识别分类,再通过Adam优化器进行细化分割。研究结果表明该模型具有一定的可行性,模型评价结果较好。输出的高精度裂缝骨架可以对缺陷类型进行判定和初步处理,为后续实际缺陷评测和修补工程提供参考。

图像级高光谱影像高分辨率特征网络分类方法

基于深度学习的高光谱影像分类方法通常将高光谱影像切分为局部方块作为模型的输入,这不但限制了长距离空-谱信息关联的获取,还带来了大量额外的计算开销。以全局图像作为输入的图像级分类方法能够有效避免这些缺陷,然而,现有的基于全卷积神经网络特征串行流动模式的图像级分类方法在信息恢复时的细节损失会导致分类精度低、分类图视觉效果差等问题。因此,本文提出一种基于HRNet的图像级高光谱影像快速分类方法,在全程保持高分辨率特征的基础上对影像的多重分辨率特征进行并行计算与交叉融合,从而缓解了传统特征串行流动模式造成的信息损失问题。同时,提出多分辨率特征联合监督和投票分类策略,进一步提升了模型分类性能。利用4组开源高光谱影像数据集对本文方法进行验证,试验结果表明,与现有的先进分类方法相比,本文方法能够取得具有竞争性的分类结果,同时显著减少训练和分类时长,在实际应用时更具时效性。为了保证方法的复现性,笔者将代码开源于https://github.com/sssssyf/fast-image-level-vote。

集成多种上下文与混合交互的显著性目标检测

显著性目标检测目的是识别和分割图像中的视觉显著性目标,它是计算机视觉任务及其相关领域的重要研究内容之一。当下基于全卷积网络(FCNs)的显著性目标检测方法已经取得了不错的性能,然而现实场景中的显著性目标类型多变且尺寸不固定,这使得准确检测并完整分割出显著性目标仍然是一个巨大的挑战。为此,该文提出集成多种上下文和混合交互的显著性目标检测方法,通过利用密集上下文信息探索模块和多源特征混合交互模块来高效预测显著性目标。密集上下文信息探索模块采用空洞卷积、不对称卷积和密集引导连接渐进地捕获具有强关联性的多尺度和多感受野上下文信息,通过集成这些信息来增强每个初始多层级特征的表达能力。多源特征混合交互模块包含多种特征聚合操作,可以自适应交互来自多层级特征中的互补性信息,以生成用于准确预测显著性图的高质量特征表示。此方法在5个公共数据集上进行了性能测试,实验结果表明,该文方法在不同的评估指标下与19种基于深度学习的显著性目标检测方法相比取得优越的预测性能。

基于深度全卷积神经弹性网络WCGAN-GP模型的语音增强研究

Wasserstein距离生成对抗网络(Wasserstein Generative Adversal Network,WGAN)模型^([1])在语音增强中运用广泛,但存在梯度易爆炸、性能不稳定等问题。引入梯度惩罚(Gradient Penalty,GP)和弹性网络条件约束,并将生成器和判别器优化成深度全卷积神经网络(Deep Fully Convolutional Neural Networks,DFCNN)结构,提出一种基于DFCNN的弹性网络条件梯度惩罚(Wasserstein Conditional Generative Adversal Network Gradient Penalty,WCGAN-GP)模型。改进后的模型可以达到真实Lipschitz限制条件,提高了可控性、稳定性和特征提取能力,能更快优化训练。实验将改进后的模型与WGAN对不同噪声条件下的语音进行增强,结果证实了改进后的模型在语音增强方面的优越性。

基于深度学习的近岸海浪图像反演有效波高算法研究

海浪有效波高是近岸海洋观测的重要要素,近岸摄像头拍摄的图像可直观地反映波高大小,但目前基于图像的有效波高反演算法研究多处于室内实验阶段且方法缺乏普适性。本文基于深度学习技术,以山东青岛小麦岛近岸海浪为例,基于海浪图像和浮标实测数据,开展近岸海浪图像反演有效波高方法研究,给出一种利用图像反演海浪有效波高的方法,该方法利用卷积网络提取海浪图像的特征,利用全连接网络提取风速等气象特征,将特征融合后作为全连接层的输入,最后输出反演的有效波高。通过对比多种模型的反演结果和浮标观测数据,发现多参数DenseNET121模型有效波高反演能力优于其他神经网络模型,其平均绝对误差为8.97 cm。本文研究为近岸海浪观测提供了一种新的技术思路。

基于轻量级全连接张量映射网络的高光谱图像分类方法

近年来,基于卷积神经网络的深度学习模型已经在高光谱图像分类领域取得优异表现.然而,模型性能的提升通常依赖于更深、更宽的网络结构,导致参数量和计算量增长,从而限制了模型在机载或星载载荷中的实际部署.为此,本文提出基于轻量级全连接张量映射网络的高光谱图像分类方法.根据全连接张量网络分解的映射思想以及高光谱图像“图谱合一”的结构特点,本文设计两种张量映射卷积单元,通过使用多个具有全连接结构的小尺寸卷积核代替原始卷积核,降低了卷积层的时间和空间复杂度.此外,基于新单元构建残差双分支张量模块.双分支结构共享同一组权重参数,并采用通道分割操作减少特征通道数,提升特征提取过程的实时性.本文所提模型通过使用新单元和新模块充分挖掘高光谱图像的局部空谱信息和全局光谱信息,有效提高了分类性能并减少硬件资源消耗.在三个常用高光谱图像数据集上的实验结果表明,所提模型相较于其他现有工作具有更高的分类性能以及更低的参数量和计算量.

一种基于全卷积神经网络的空中目标战术意图识别模型

针对现有空中目标识别方法敏捷性和可靠度不够高的问题,研究设计了一种深度学习模型MLSTM-FCN,结合了全卷积神经网络、循环神经网络和压缩与激励模块的优点。全卷积网络能够提取空战数据中的复杂局部特征,长短记忆神经网络可以捕捉空战意图数据的时序特征。通过消融实验和对比实验结果表明,MLSTM-FCN模型在意图识别准确率、反应速度和抗干扰能力方面明显优于现有的空中目标意图识别模型,取得了sota的结果,为指挥员在进行空中作战决策时提供更有效的依据。

嵌入NLB模块的FCN在轴承信号降噪中的应用

深度学习在故障诊断取得了显著的进展,然而其多为端到端的智能诊断,在信号降噪方面的应用较少。本文提出了一种基于全卷积神经网络(fully convolutional network,FCN)的降噪方法。首先,模型整体采用了encoder-decoder架构,其中encoder部分由三层卷积层组成,decoder部分由四层反卷积层组成。其次,引入了残差连接对模型的学习目标进行了约束,使得模型在传播过程中更多地关注噪声信息。并且为了增强模型的特征提取能力,在encoder和decoder中引入了非局部块(non-local block,NLB)。然后,通过仿真信号对比实验选择网络的超参数,与目前主流的降噪方法进行对比,初步验证模型的降噪效果。最后,通过实际案例对所提方法的降噪效果进行对比验证,结果表明本文提出的方法在直观观察和降噪性能指标方面均取得了良好的应用效果,能够有效提高故障诊断的准确率。

基于全卷积编码-解码对称网络的单目图像深度估计

单目图像深度估计通过唯一视角下的图像来感知每个像素的空间位置关系,对于场景理解、三维重建等具有重要意义。为了全面提升预测深度图涵盖的信息量,保持关键细节不丢失,论文基于对称的编解码结构设计了一个全卷积网络来执行深度估计任务,称为ResUNet。该网络继承了U-Net模型的经典架构,首先采用了改进的ResNet网络来实现特征编码,其次保留了U-Net模型的解码器设计来将特征图解码为深度图,这种结构设计融合了ResNet和U-Net网络的特点,通过协同优化最大程度发挥了各自的优势,能够在进行深度估计的过程中实现空间结构和细节信息的最大程度保留,进而提升预测深度图的真实性与可靠性。基于该网络进一步提出了ResDepth算法,该算法针对深度图预测过程中容易产生物体失真、细节淹没的问题,从损失函数的角度出发设计了一个联合损失函数,在不带来额外计算开销的情况下全面提升了预测深度图的质量。最后,在NYU-Depth V2、SUN RGB-D及KITTI三个公开数据集上进行对比实验来评估算法性能,实验表明,论文提出的ResDepth算法及联合损失函数能够更好地保留空间结构信息及几何细节信息,进而提升深度估计结果的准确性。

改进FCOS算法的车辆检测方法研究

针对目前车辆检测的误差率高、检测速度慢等问题,提出一种基于改进全卷积单阶段(Fully Convolutional One-Stage Object Detection,FCOS)的车辆检测算法。通过引入一种考虑多个几何特征的交并比损失函数,改善了训练过程中高长宽比车辆、并行车辆难以准确回归的现象;使用多尺度卷积结合多维特征信息,增强了算法对不同尺度检测的鲁棒性;根据车辆检测场景改进了回归尺度,提高模型的推理准确度。实验结果表明,该方法在车辆检测任务中能够明显提升检测精度并保持检测速度不下降。