Diagnosis of Middle Ear Diseases Based on Convolutional Neural Network

An otoscope is traditionally used to examine the eardrum and ear canal.A diagnosis of otitis media(OM)relies on the experience of clinicians.If an examiner lacks experience,the examination may be difficult and time-consuming.This paper presents an ear disease classification method using middle ear images based on a convolutional neural network(CNN).Especially the segmentation and classification networks are used to classify an otoscopic image into six classes:normal,acute otitis media(AOM),otitis media with effusion(OME),chronic otitis media(COM),congenital cholesteatoma(CC)and traumatic perforations(TMPs).The Mask R-CNN is utilized for the segmentation network to extract the region of interest(ROI)from otoscopic images.The extracted ROIs are used as guiding features for the classification.The classification is based on transfer learning with an ensemble of two CNN classifiers:EfficientNetB0 and Inception-V3.The proposed model was trained with a 5-fold cross-validation technique.The proposed method was evaluated and achieved a classification accuracy of 97.29%.

DNN-Based Speech Enhancement Using Soft Audible Noise Masking for Wind Noise Reduction

This paper presents a deep neural network(DNN)-based speech enhancement algorithm based on the soft audible noise masking for the single-channel wind noise reduction. To reduce the low-frequency residual noise, the psychoacoustic model is adopted to calculate the masking threshold from the estimated clean speech spectrum. The gain for noise suppression is obtained based on soft audible noise masking by comparing the estimated wind noise spectrum with the masking threshold. To deal with the abruptly time-varying noisy signals, two separate DNN models are utilized to estimate the spectra of clean speech and wind noise components. Experimental results on the subjective and objective quality tests show that the proposed algorithm achieves the better performance compared with the conventional DNN-based wind noise reduction method.

A Deep Learning Model of Traffic Signs in Panoramic Images Detection

To pursue the ideal of a safe high-tech society in a time when traffic accidents are frequent,the traffic signs detection system has become one of the necessary topics in recent years and in the future.The ultimate goal of this research is to identify and classify the types of traffic signs in a panoramic image.To accomplish this goal,the paper proposes a new model for traffic sign detection based on the Convolutional Neural Network for com-prehensive traffic sign classification and Mask Region-based Convolutional Neural Networks(R-CNN)implementation for identifying and extracting signs in panoramic images.Data augmentation and normalization of the images are also applied to assist in classifying better even if old traffic signs are degraded,and considerably minimize the rates of discovering the extra boxes.The proposed model is tested on both the testing dataset and the actual images and gets 94.5%of the correct signs recognition rate,the classification rate of those signs discovered was 99.41%and the rate of false signs was only around 0.11.

基于掩码提示与门控记忆网络校准的关系抽取方法

针对关系抽取(RE)任务中实体关系语义挖掘困难和预测关系有偏差等问题,提出一种基于掩码提示与门控记忆网络校准(MGMNC)的RE方法。首先,利用提示中的掩码学习实体之间在预训练语言模型(PLM)语义空间中的潜在语义,通过构造掩码注意力权重矩阵,将离散的掩码语义空间相互关联;其次,采用门控校准网络将含有实体和关系语义的掩码表示融入句子的全局语义;再次,将它们作为关系提示校准关系信息,随后将句子表示的最终表示映射至相应的关系类别;最后,通过更好地利用提示中掩码,并结合传统微调方法的学习句子全局语义的优势,充分激发PLM的潜力。实验结果表明,所提方法在SemEval(SemEval-2010 Task 8)数据集的F1值达到91.4%,相较于RELA(Relation Extraction with Label Augmentation)生成式方法提高了1.0个百分点;在SciERC(Entities, Relations, and Coreference for Scientific knowledge graph construction)和CLTC(Chinese Literature Text Corpus)数据集上的F1值分别达到91.0%和82.8%。所提方法在上述3个数据集上均明显优于对比方法,验证了所提方法的有效性。相较于基于生成式的方法,所提方法实现了更优的抽取性能。

基于掩码语言模型的中文BERT攻击方法

对抗文本是一种能够使深度学习分类器作出错误判断的恶意样本,敌手通过向原始文本中加入人类难以察觉的微小扰动制作出能欺骗目标模型的对抗文本.研究对抗文本生成方法,能对深度神经网络的鲁棒性进行评价,并助力于模型后续的鲁棒性提升工作.当前针对中文文本设计的对抗文本生成方法中,很少有方法将鲁棒性较强的中文BERT模型作为目标模型进行攻击.面向中文文本分类任务,提出一种针对中文BERT的攻击方法Chinese BERT Tricker.该方法使用一种汉字级词语重要性打分方法——重要汉字定位法;同时基于掩码语言模型设计一种包含两类策略的适用于中文的词语级扰动方法实现对重要词语的替换.实验表明,针对文本分类任务,所提方法在两个真实数据集上均能使中文BERT模型的分类准确率大幅下降至40%以下,且其多种攻击性能明显强于其他基线方法.

刮板机异常监测系统设计

为了实时识别刮板机上的异常小目标,确保刮板机的正常、安全运行,设计了基于机器视觉的刮板机异常监测系统。数据采集层的工业摄像机采集单元基于机器视觉原理获取刮板机实时监测图像,经通用串行总线(USB)接口传输图像给数据处理层。对采集的刮板机图像作降噪、增强处理后,通过数据传输层的基于现场可编程门阵列(FPGA)的以太网通信模块完成图像的上传。数据监测层的异常状态监测模块依据接收到的图像,创新性地调用改进的掩蔽区域卷积神经网络(Mask R-CNN)模型,由异常报警模块发送报警信息,并通过数据显示层呈现异常监测结果及报警提示信息,以实现刮板机异常监测。试验结果表明:该系统处理后的刮板机图像峰值信噪比显著提升、均方根误差显著降低;增强后的刮板机图像异常识别损失更低。该系统可识别刮板机不同类型的异常,并标记异常目标。

Mask R-CNN and multifeature clustering model for catenary insulator recognition and defect detection

Rod insulators are vital parts of the catenary of high speed railways(HSRs).There are many different catenary insulators,and the background of the insulator image is complicated.It is difficult to recognise insulators and detect defects automatically.In this paper,we propose a catenary intelligent defect detection algorithm based on Mask region-convolutional neural network(R-CNN)and an image processing model.Vertical projection technology is used to achieve single shed positioning and precise cutting of the insulator.Gradient,texture,and gray feature fusion(GTGFF)and a K-means clustering analysis model(KCAM)are proposed to detect broken insulators,dirt,foreign bodies,and flashover.Using this model,insulator recognition and defect detection can achieve a high recall rate and accuracy,and generalized defect detection.The algorithm is tested and verified on a dataset of realistic insulator images,and the accuracy and reliability of the algorithm satisfy current requirements for HSR catenary automatic inspection and intelligent maintenance.

基于注意力机制的稀疏化剪枝方法

为在资源受限设备中部署先进神经网络模型,提出一种基于通道和空间注意力机制的网络稀疏化剪枝训练方法,将剪枝训练过程转化为约束优化问题。将通道和空间注意力融入稀疏化剪枝训练过程,利用连续空间损失变化情况评估不同网络层重要程度,通过稀疏化训练与动态计算及更新掩码矩阵和权重矩阵完成剪枝操作。方法实验基于CIFAR10、CIFAR100数据集上进行,实验结果表明,该方法在较为复杂数据集CIFAR100上剪枝率为90%、95%、98%时,分类准确率可达到69.91%、67.15%、60.18%,与同类方法相比,在不同数据集和剪枝率的条件下仍具有较高的分类精度。

基于智能破碎锤作业场景的多目标石块识别定位方法

随着智能工程机械的不断发展,结构复杂、功能单一的传统液压破碎锤逐渐被日益成熟的智能破碎锤取代,对多目标石块的识别与定位是智能破碎提高动作输出精度、完成破碎任务的重要保证.提出基于分割掩码卷积神经网络(Mask R-CNN)实例分割和激光雷达信息融合的目标石块识别定位方法,通过Mask R-CNN实例分割算法快速识别复杂作业场景下目标石块的感兴趣区(Region of Interest,RoI);在保证石块检测精确率的前提下,融合激光雷达通过卡尔曼滤波算法得到破碎点位置信息,引导破碎锤实现定位作业.现场试验结果表明,目标石块检测模型对石块的平均识别精确率为95.35%,召回率为95.06%,石块破碎点识别精确率为94.20%.在复杂作业背景下,该方法可实现多目标石块识别和破碎点定位,满足自动破碎实时性要求.

基于YOLO与改进的DLIB多角度遮挡人脸判别方法

针对ATM机上违法犯罪分子通过遮挡面部进行犯罪活动进而无法追踪的问题,提出了一种基于YOLO与改进的DLIB多角度遮挡人脸判别方法。通过将基于YOLO模型的多目标检测改成单一人脸检测,并调整其损失函数中人脸置信度损失计算方式,提高了人脸定位的准确性与时效性,完成了从原始图像的输入到任意人脸位置的回归,再结合改进的DLIB多角度人脸68个关键点检测算法在回归出的人脸位置上进行遮挡判别的新方法。测试结果验证了新方法优于传统方法,能够有效并快速地判别出各类遮挡,实现了ATM机上遮挡人脸判别的实时性与鲁棒性,具有较高的应用价值。

Localization and Classification of Rice-grain Images Using Region Proposals-based Convolutional Neural Network

This paper proposes a solution to localization and classification of rice grains in an image.All existing related works rely on conventional based machine learning approaches.However,those techniques do not do well for the problem designed in this paper,due to the high similarities between different types of rice grains.The deep learning based solution is developed in the proposed solution.It contains pre-processing steps of data annotation using the watershed algorithm,auto-alignment using the major axis orientation,and image enhancement using the contrast-limited adaptive histogram equalization(CLAHE)technique.Then,the mask region-based convolutional neural networks(R-CNN)is trained to localize and classify rice grains in an input image.The performance is enhanced by using the transfer learning and the dropout regularization for overfitting prevention.The proposed method is validated using many scenarios of experiments,reported in the forms of mean average precision(mAP)and a confusion matrix.It achieves above 80%mAP for main scenarios in the experiments.It is also shown to perform outstanding,when compared to human experts.

口罩佩戴识别中的Tiny-YOLOv3模型算法优化

针对深度学习网络(Tiny-YOLOv3)算法准确率不高以及更改网络模型后实时性的问题,提出一种网络改进方案和基于BN层剪枝的优化算法。将Tiny-YOLOv3的前四层池化层改为两步长的卷积层进行下采样以及增加特征的提取,将后两层池化层和第六个卷积层改为一个残差结构层,再利用BN层剪枝算法,将网络进行压缩和BN层合并来加速网络。改进优化后的模型算法相比原始Tiny-YOLOv3网络,在口罩佩戴识别的平均精确率(mAP)提升了14%,模型体积只有19.2 MB,压缩了42%;平均每秒传输帧数(FPS)增加了17%。实验结果表明,改进优化后的模型有更好的精确性和实时性。

面曝光快速成形系统制件的强度模型研究

为了快速准确地获得面曝光快速成形制件的强度,利用传统多项式和BP神经网络建立了制件强度模型,建模结果显示二次多项式模型的最大偏差为9.5632 MPa,平均偏差为2.3812 MPa,BP神经网络模型的最大偏差为4.997 MPa,平均偏差为0.8435 MPa。研究结果表明:BP神经网络模型计算结果优于二次多项式模型,并具有一定的预测能力,可用于面曝光快速成形系统制件强度模型的建立。

基于像素偏转模型和机器学习的室外图像天空像素检测

图像中的天空区域对于基于视觉的地面机器人导航具有重要意义,为了识别图像中的天空部分,本文提出了一种基于像素偏转模型的BP神经网络天空识别方法。首先,制作天空图像集和非天空图像集,天空图像集由各种天气情况下的天空提取而成,非天空图像集由非天空的景物构成,主要包括建筑、汽车、树木、植物等;其次,使用提出的像素偏转模型提取天空图像集和非天空图像集的像素特征并进行处理,对天空像素点和非天空像素点进行标注,利用BP神经网络对像素特征进行训练,得到权重文件;最后,使用得到的权重文件进行天空的识别。为了更好的说明本文算法和模型的优越性,使用本文算法与Otsu算法、Ye Hu算法、Graph-cut算法和Mask-Rcnn算法模型进行了比较,并设计了两组组对比实验,第一组实验进行识别效果的主观评价,第二组实验利用Cam Vid数据集的天空类进行算法精度的定量分析。

基于无人机图像的小麦麦穗自动识别方法研究

小麦麦穗智能识别在产量预估、种子筛选、农田密度估计等方面具有一定的科研价值和应用价值。传统的麦穗识别方法效率低且浪费人力,因此提出以无人机高清图像为数据源,基于卷积神经网络建立Mask-RCNN深度神经网络的小麦麦穗自动识别模型,以平均精准均值(Mean Average Precision,MAP)和预测准确率(Precision)作为模型的评价指标。实验结果表明,Mask-RCNN麦穗识别模型的MAP达到72.32%,准确率达到95.2%,说明基于Mask-RCNN的麦穗识别模型是一种高效准确的识别方法,具有一定的实用价值。

基于BP神经网络的口罩防护舒适性研究

冬季雾霾天气频现，然而可供选择的口罩很少。探讨了以客观指标作为输入参数的口罩防护舒适性的预测结果。通过测试16种口罩试样的5个客观评价指标以及4个主观防护舒适性评价指标，建立预测口罩防护舒适性的BP神经网络模型。以客观测试指标为模型的输入参数，以主观评价指标为模型的输出参数，并在16组试样中随机选取4组来判断模型的准确性。结论显示：模型对防霾口罩的防护舒适性预测具有一定的可靠性，其预测绝对误差范围在0．5—0．8。