深度卷积神经网络应用于人脸特征点检测研究

为解决在复杂环境下,如姿势不同、光照条件以及遮挡等因素导致传统人脸特征点检测算法的精度大幅度下降的问题,在特征点检测理论知识以及研究现状的基础上,针对传统卷积神经网络模型在处理人脸特征点检测问题时的不足之处,提出基于小滤波器的深卷积神经网络。算法引入小滤波器思想和以拓展"网络深度"优先的深层卷积神经网络模型,针对人脸特征点检测重新设计训练,提高了算法的有效性与适用性。通过将算法应用于ALFW和AFW人脸数据集上预测5点人脸特征点问题,并与其他多个经典算法进行对比分析,结果表明:基于小滤波器的深卷积神经网络在预测人脸5点特征点问题上有更好的准确性和鲁棒性。

基于卷积神经网络的步态周期检测方法

针对步态周期检测直接影响到步态识别的计算量和精度的问题,本文基于深卷积神经网络实现步态周期检测,分别通过将步态序列根据周期性进行分类,和将步态序列拟合为正弦函数进行步态周期性检测的方法,对步态周期进行建模。其关键思想是根据步态周期的规律性,将步态起伏作为分类问题或一个正弦函数来建模,步态视频中的每个帧对应一个可以表示其周期特征的类别或函数值。而卷积神经网络被用于提取步态帧的周期性特征,以定位该帧在周期中的位置,得到分类或回归结果,最终实现步态周期检测。在CASIA-B数据集的不同视角下采用了多种网络结构进行了周期检测效果的验证。实验结果表明:该方法对步态周期性检测具有良好的精度和鲁棒性。

一种用于时空图像融合的增强深卷积模型

研究一种基于增强深卷积模型的时空图像融合算法。设置在多个视频探头和激光点云探头的数据支持下,使用四个增强卷积神经网络模块进行数据预处理,并将预处理数据使用多列神经网络整理,直接控制两侧履带和探头云台。当该神经网络训练时间超过150h时,机器人视觉自主驱动的行走运行效率达到工业化运行要求。

Rapid urban flood forecasting based on cellular automata and deep learning

[Objective]Urban floods are occurring more frequently because of global climate change and urbanization.Accordingly,urban rainstorm and flood forecasting has become a priority in urban hydrology research.However,two-dimensional hydrodynamic models execute calculations slowly,hindering the rapid simulation and forecasting of urban floods.To overcome this limitation and accelerate the speed and improve the accuracy of urban flood simulations and forecasting,numerical simulations and deep learning were combined to develop a more effective urban flood forecasting method.[Methods]Specifically,a cellular automata model was used to simulate the urban flood process and address the need to include a large number of datasets in the deep learning process.Meanwhile,to shorten the time required for urban flood forecasting,a convolutional neural network model was used to establish the mapping relationship between rainfall and inundation depth.[Results]The results show that the relative error of forecasting the maximum inundation depth in flood-prone locations is less than 10%,and the Nash efficiency coefficient of forecasting inundation depth series in flood-prone locations is greater than 0.75.[Conclusion]The result demonstrated that the proposed method could execute highly accurate simulations and quickly produce forecasts,illustrating its superiority as an urban flood forecasting technique.

第三届胶囊内镜全球高峰论坛纪要

为关注胶囊内镜的消化检查防疫,展现2020年胶囊内镜最新发展成果,同时加强国际领域在胶囊内镜产品、质控、研发、临床应用、人才的交流与合作,2020第三届胶囊内镜全球高峰论坛在重庆召开.峰会邀请国外专家线上远程直播专题学术演讲,受邀的国内专家现场带来最新学术报告,共有17位医学专家从5个层面呈现了胶囊内镜领域的多项最新技术和学术成果.世界华人消化学会会长、《世界华人消化杂志》主编江学良教授应邀作了胶囊内镜临床应用的报告.

基于ERNIE-DPCNN的短文本分类研究

目前采用短文本分类的方法几乎都使用词向量,不管是机器学习还是深度学习本质上都是对数字的处理。将文本汉字转换成计算机可识别的数字信息是词向量的作用。ERNIE是百度提出主要针对中文设计的词向量模型。将ERNIE词向量与深金字塔卷积神经网络相融合,对中文类新闻文本标题进行文本分类处理。通过实验比较,ERNIE词向量与深金字塔卷积神经网络相结合的短文本分类模型具有较高的分类精度。

Neighborhood fusion-based hierarchical parallel feature pyramid network for object detection

In order to improve the detection accuracy of small objects,a neighborhood fusion-based hierarchical parallel feature pyramid network(NFPN)is proposed.Unlike the layer-by-layer structure adopted in the feature pyramid network(FPN)and deconvolutional single shot detector(DSSD),where the bottom layer of the feature pyramid network relies on the top layer,NFPN builds the feature pyramid network with no connections between the upper and lower layers.That is,it only fuses shallow features on similar scales.NFPN is highly portable and can be embedded in many models to further boost performance.Extensive experiments on PASCAL VOC 2007,2012,and COCO datasets demonstrate that the NFPN-based SSD without intricate tricks can exceed the DSSD model in terms of detection accuracy and inference speed,especially for small objects,e.g.,4%to 5%higher mAP(mean average precision)than SSD,and 2%to 3%higher mAP than DSSD.On VOC 2007 test set,the NFPN-based SSD with 300×300 input reaches 79.4%mAP at 34.6 frame/s,and the mAP can raise to 82.9%after using the multi-scale testing strategy.

Prediction of Departure Aircraft Taxi Time Based on Deep Learning

With the continuous increase in the number of flights,the use of airport collaborative decision-making(ACDM)systems has been more and more widely spread.The accuracy of the taxi time prediction has an important effect on the A-CDM calculation of the departure aircraft’s take-off queue and the accurate time for the aircraft blockout.The spatial-temporal-environment deep learning(STEDL)model is presented to improve the prediction accuracy of departure aircraft taxi-out time.The model is composed of time-flow sub-model(airport capacity,number of taxiing aircraft,and different time periods),spatial sub-model(taxiing distance)and environmental sub-model(weather,air traffic control,runway configuration,and aircraft category).The STEDL model is used to predict the taxi time of departure aircraft at Hong Kong Airport and the results show that the STEDL method has a prediction accuracy of 95.4%.The proposed model also greatly reduces the prediction error rate compared with the other machine learning methods.

Tongue image segmentation and tongue color classification based on deep learning

Objective To propose two novel methods based on deep learning for computer-aided tongue diagnosis,including tongue image segmentation and tongue color classification,improving their diagnostic accuracy.Methods LabelMe was used to label the tongue mask and Snake model to optimize the labeling results.A new dataset was constructed for tongue image segmentation.Tongue color was marked to build a classified dataset for network training.In this research,the Inception+Atrous Spatial Pyramid Pooling(ASPP)+UNet(IAUNet)method was proposed for tongue image segmentation,based on the existing UNet,Inception,and atrous convolution.Moreover,the Tongue Color Classification Net(TCCNet)was constructed with reference to ResNet,Inception,and Triple-Loss.Several important measurement indexes were selected to evaluate and compare the effects of the novel and existing methods for tongue segmentation and tongue color classification.IAUNet was compared with existing mainstream methods such as UNet and DeepLabV3+for tongue segmentation.TCCNet for tongue color classification was compared with VGG16 and GoogLeNet.Results IAUNet can accurately segment the tongue from original images.The results showed that the Mean Intersection over Union(MIoU)of IAUNet reached 96.30%,and its Mean Pixel Accuracy(MPA),mean Average Precision(mAP),F1-Score,G-Score,and Area Under Curve(AUC)reached 97.86%,99.18%,96.71%,96.82%,and 99.71%,respectively,suggesting IAUNet produced better segmentation than other methods,with fewer parameters.Triplet-Loss was applied in the proposed TCCNet to separate different embedded colors.The experiment yielded ideal results,with F1-Score and mAP of the TCCNet reached 88.86% and 93.49%,respectively.Conclusion IAUNet based on deep learning for tongue segmentation is better than traditional ones.IAUNet can not only produce ideal tongue segmentation,but have better effects than those of PSPNet,SegNet,UNet,and DeepLabV3+,the traditional networks.As for tongue color classification,the proposed network,TCCNet,had better F1-Score and mAP values as compared with other neural networks such as VGG16 and GoogLeNet.

External and Internal Validation of a Computer Assisted Diagnostic Model for Detecting Multi-Organ Mass Lesions in CT images

Objective We developed a universal lesion detector(ULDor)which showed good performance in in-lab experiments.The study aims to evaluate the performance and its ability to generalize in clinical setting via both external and internal validation.Methods The ULDor system consists of a convolutional neural network(CNN)trained on around 80 K lesion annotations from about 12 K CT studies in the DeepLesion dataset and 5 other public organ-specific datasets.During the validation process,the test sets include two parts:the external validation dataset which was comprised of 164 sets of non-contrasted chest and upper abdomen CT scans from a comprehensive hospital,and the internal validation dataset which was comprised of 187 sets of low-dose helical CT scans from the National Lung Screening Trial(NLST).We ran the model on the two test sets to output lesion detection.Three board-certified radiologists read the CT scans and verified the detection results of ULDor.We used positive predictive value(PPV)and sensitivity to evaluate the performance of the model in detecting space-occupying lesions at all extra-pulmonary organs visualized on CT images,including liver,kidney,pancreas,adrenal,spleen,esophagus,thyroid,lymph nodes,body wall,thoracic spine,etc.Results In the external validation,the lesion-level PPV and sensitivity of the model were 57.9%and 67.0%,respectively.On average,the model detected 2.1 findings per set,and among them,0.9 were false positives.ULDor worked well for detecting liver lesions,with a PPV of 78.9%and a sensitivity of 92.7%,followed by kidney,with a PPV of 70.0%and a sensitivity of 58.3%.In internal validation with NLST test set,ULDor obtained a PPV of 75.3%and a sensitivity of 52.0%despite the relatively high noise level of soft tissue on images.Conclusions The performance tests of ULDor with the external real-world data have shown its high effectiveness in multiple-purposed detection for lesions in certain organs.With further optimisation and iterative upgrades,ULDor may be well suited for extensive application to external data.

Object detection of artifact threaded hole based on Faster R-CNN

In order to improve the accuracy of threaded hole object detection,combining a dual camera vision system with the Hough transform circle detection,we propose an object detection method of artifact threaded hole based on Faster region-ased convolutional neural network(Faster R-CNN).First,a dual camera image acquisition system is established.One industrial camera placed at a high position is responsible for collecting the whole image of the workpiece,and the suspected screw hole position on the workpiece can be preliminarily selected by Hough transform detection algorithm.Then,the other industrial camera is responsible for collecting the local images of the suspected screw holes that have been detected by Hough transform one by one.After that,ResNet50-based Faster R-CNN object detection model is trained on the self-built screw hole data set.Finally,the local image of the threaded hole is input into the trained Faster R-CNN object detection model for further identification and location.The experimental results show that the proposed method can effectively avoid small object detection of threaded holes,and compared with the method that only uses Hough transform or Faster RCNN object detection alone,it has high recognition and positioning accuracy.

A deep dense captioning framework with joint localization and contextual reasoning

Dense captioning aims to simultaneously localize and describe regions-of-interest(RoIs)in images in natural language.Specifically,we identify three key problems:1)dense and highly overlapping RoIs,making accurate localization of each target region challenging;2)some visually ambiguous target regions which are hard to recognize each of them just by appearance;3)an extremely deep image representation which is of central importance for visual recognition.To tackle these three challenges,we propose a novel end-to-end dense captioning framework consisting of a joint localization module,a contextual reasoning module and a deep convolutional neural network(CNN).We also evaluate five deep CNN structures to explore the benefits of each.Extensive experiments on visual genome(VG)dataset demonstrate the effectiveness of our approach,which compares favorably with the state-of-the-art methods.

Arrival Pattern Recognition and Prediction Based on Machine Learning

A data-driven method for arrival pattern recognition and prediction is proposed to provide air traffic controllers(ATCOs)with decision support. For arrival pattern recognition,a clustering-based method is proposed to cluster arrival patterns by control intentions. For arrival pattern prediction,two predictors are trained to estimate the most possible command issued by the ATCOs in a particular traffic situation. Training the arrival pattern predictor could be regarded as building an ATCOs simulator. The simulator can assign an appropriate arrival pattern for each arrival aircraft,just like real ATCOs do. Therefore,the simulator is considered to be able to provide effective advice for part of the work of ATCOs. Finally,a case study is carried out and demonstrates that the convolutional neural network(CNN)-based predictor performs better than the radom forest(RF)-based one.

Research on single image super-resolution based on very deep super-resolution convolutional neural network

Single image super-resolution(SISR)is a fundamentally challenging problem because a low-resolution(LR)image can correspond to a set of high-resolution(HR)images,while most are not expected.Recently,SISR can be achieved by a deep learning-based method.By constructing a very deep super-resolution convolutional neural network(VDSRCNN),the LR images can be improved to HR images.This study mainly achieves two objectives:image super-resolution(ISR)and deblurring the image from VDSRCNN.Firstly,by analyzing ISR,we modify different training parameters to test the performance of VDSRCNN.Secondly,we add the motion blurred images to the training set to optimize the performance of VDSRCNN.Finally,we use image quality indexes to evaluate the difference between the images from classical methods and VDSRCNN.The results indicate that the VDSRCNN performs better in generating HR images from LR images using the optimized VDSRCNN in a proper method.

Classification of hyperspectral images based on a convolutional neural network and spectral sensitivity

In recent years,deep learning methods have gradually come to be used in hyperspectral imaging domains.Because of the peculiarity of hyperspectral imaging,a mass of information is contained in the spectral dimensions of hyperspectral images.Also,different ob jects on a land surface are sensitive to different ranges of wavelength.To achieve higher accuracy in classification,we propose a structure that combines spectral sensitivity with a convolutional neural network by adding spectral weights derived from predicted outcomes before the final classification layer.First,samples are divided into visible light and infrared,with a portion of the samples fed into networks during training.Then,two key parameters,unrecognized rate(δ)and wrongly recognized rate(γ),are calculated from the predicted outcome of the whole scene.Next,the spectral weight,derived from these two parameters,is calculated.Finally,the spectral weight is added and an improved structure is constructed.The improved structure not only combines the features in spatial and spectral dimensions,but also gives spectral sensitivity a primary status.Compared with inputs from the whole spectrum,the improved structure attains a nearly 2%higher prediction accuracy.When applied to public data sets,compared with the whole spectrum,on the average we achieve approximately 1%higher accuracy.

Automatic anatomical classification of colonoscopic images using deep convolutional neural networks

Background:A colonoscopy can detect colorectal diseases,including cancers,polyps,and inflammatory bowel diseases.A computer-aided diagnosis(CAD)system using deep convolutional neural networks(CNNs)that can recognize anatomical locations during a colonoscopy could efficiently assist practitioners.We aimed to construct a CAD system using a CNN to distinguish colorectal images from parts of the cecum,ascending colon,transverse colon,descending colon,sigmoid colon,and rectum.Method:We constructed a CNN by training of 9,995 colonoscopy images and tested its performance by 5,121 independent colonoscopy images that were categorized according to seven anatomical locations:the terminal ileum,the cecum,ascending colon to transverse colon,descending colon to sigmoid colon,the rectum,the anus,and indistinguishable parts.We examined images taken during total colonoscopy performed between January 2017 and November 2017 at a single center.We evaluated the concordance between the diagnosis by endoscopists and those by the CNN.The main outcomes of the study were the sensitivity and specificity of the CNN for the anatomical categorization of colonoscopy images.Results:The constructed CNN recognized anatomical locations of colonoscopy images with the following areas under the curves:0.979 for the terminal ileum;0.940 for the cecum;0.875 for ascending colon to transverse colon;0.846 for descending colon to sigmoid colon;0.835 for the rectum;and 0.992 for the anus.During the test process,the CNN system correctly recognized 66.6%of images.Conclusion:We constructed the new CNN system with clinically relevant performance for recognizing anatomical locations of colonoscopy images,which is the first step in constructing a CAD system that will support us during colonoscopy and provide an assurance of the quality of the colonoscopy procedure.