微动脉瘤是糖尿病视网膜病变的初期症状,消除该病灶可在早期非常有效地预防糖尿病视网膜病变。但由于视网膜结构复杂,同时眼底图像的成像由于患者、环境、采集设备等因素的不同会存在不同的亮度和对比度,现有的微动脉瘤检测算法难以实...微动脉瘤是糖尿病视网膜病变的初期症状,消除该病灶可在早期非常有效地预防糖尿病视网膜病变。但由于视网膜结构复杂,同时眼底图像的成像由于患者、环境、采集设备等因素的不同会存在不同的亮度和对比度,现有的微动脉瘤检测算法难以实现该病灶的精确检测和定位,为此本文提出嵌入SENet(squeeze-andexcitation networks)的改进YOLO(you only look once)v4自动检测算法。该算法在YOLOv4网络基础上,首先通过使用一种改进的快速模糊C均值聚类算法对目标样本进行先验框参数优化,以提高先验框与特征图的匹配度;然后,在主干网络嵌入SENet模块,通过强化关键信息,抑制背景信息,提高微动脉瘤的置信度;此外,还在网络颈部增加空间金字塔池化结构以增强主干网络输出特征的接受域,从而有助于分离出重要的上下文信息;最后,在Kaggle数据集上进行模型验证,并与其他方法进行对比。实验结果表明,与其他各种结构的YOLOv4网络模型相比,所提出的嵌入SENet的改进YOLOv4网络模型能显著提高检测结果(与原始YOLOv4相比Fscore提升了12.68%);与其他网络模型以及方法相比,所提出的嵌入SENet的改进YOLOv4网络模型的自动检测精度明显更优,且可实现精准定位。故本文所提出的嵌入SENet的改进YOLOv4算法性能较优,能准确、有效地检测并定位出眼底图像中的微动脉瘤。展开更多
针对无人机航拍图像目标检测中视野变化大、时空信息复杂等问题,文中基于YOLOv5(You Only Look Once Version5)架构,提出基于图像低维特征融合的航拍小目标检测模型.引入CA(Coordinate Attention),改进MobileNetV3的反转残差块,增加图...针对无人机航拍图像目标检测中视野变化大、时空信息复杂等问题,文中基于YOLOv5(You Only Look Once Version5)架构,提出基于图像低维特征融合的航拍小目标检测模型.引入CA(Coordinate Attention),改进MobileNetV3的反转残差块,增加图像空间维度信息的同时降低模型参数量.改进YOLOv5特征金字塔网络结构,融合浅层网络中的特征图,增加模型对图像低维有效信息的表达能力,进而提升小目标检测精度.同时为了降低航拍图像中复杂背景带来的干扰,引入无参平均注意力模块,同时关注图像的空间注意力与通道注意力;引入VariFocal Loss,降低负样本在训练过程中的权重占比.在VisDrone数据集上的实验验证文中模型的有效性,该模型在有效提升检测精度的同时明显降低复杂度.展开更多
In the experiment, the brown rice whose moisture content was 12.5% was used as raw material. The brown rice was grouped, then moisturized differently and milled. While milling, the energy consumption, the rate of brok...In the experiment, the brown rice whose moisture content was 12.5% was used as raw material. The brown rice was grouped, then moisturized differently and milled. While milling, the energy consumption, the rate of broken rice and the crack rate were tested. It is confirmed that the stress crack owing to the moisture added to the brown rice can be avoided when the moisture amount added once is limited to no more than 1.5%. It is also proved that the energy consumption can be reduced, the yielding rate of rice can be increased and that the quality of rice can be improved.展开更多
目标检测是计算机视觉领域的研究热点和基础任务,其中基于锚点(Anchor)的目标检测已在众多领域得到广泛应用。当前锚点选取方法主要面临两个问题:基于特定数据集的先验取值尺寸固定、面对不同场景泛化能力弱。计算锚框的无监督K-means算...目标检测是计算机视觉领域的研究热点和基础任务,其中基于锚点(Anchor)的目标检测已在众多领域得到广泛应用。当前锚点选取方法主要面临两个问题:基于特定数据集的先验取值尺寸固定、面对不同场景泛化能力弱。计算锚框的无监督K-means算法,受初始值影响较大,对目标尺寸较单一的数据集聚类产生的锚点差异较小,无法充分体现网络多尺度输出的特点。针对上述问题,本文提出一种基于多尺度的目标检测锚点构造方法(multi-scale-anchor,MSA),将聚类产生的锚点根据数据集本身的特性进行尺度的缩放和拉伸,优化的锚点即保留原数据集的特点也体现了模型多尺度的优势。另外,本方法应用在训练的预处理阶段,不增加模型推理时间。最后,选取单阶段主流算法YOLO(You Only Look Once),在多个不同场景的红外或工业场景数据集上进行丰富的实验。结果表明,多尺度锚点优化方法MSA能显著提高小样本场景的检测精度。展开更多
针对部署于有限算力平台的YOLOv3(you only look once v3)算法对电容器外观缺陷存在检测速度较慢的问题,提出了基于YOLOv3算法改进的轻量化算法MQYOLOv3。首先采用轻量化网络MobileNet v2作为特征提取模块,通过利用深度可分离式卷积替...针对部署于有限算力平台的YOLOv3(you only look once v3)算法对电容器外观缺陷存在检测速度较慢的问题,提出了基于YOLOv3算法改进的轻量化算法MQYOLOv3。首先采用轻量化网络MobileNet v2作为特征提取模块,通过利用深度可分离式卷积替换一般卷积操作,使得模型的参数量大幅度降低进而提高模型的检测速度,同时也带来了检测精度的降低;然后在网络结构中嵌入空间金字塔池化结构实现局部特征与全局特征的融合、引入距离交并比(distance intersection over union,DIoU)损失函数优化交并比(intersection over union,IoU)损失函数以及使用Mish激活函数优化Leaky ReLU激活函数来提高模型的检测精度。本文采用自制的电容器外观缺陷数据集进行实验,轻量化MQYOLOv3算法的平均精度均值(mean average precision,mAP)为87.96%,较优化前降低了1.16%,检测速度从1.5 FPS提升到7.7 FPS。实验表明,本文设计的轻量化MQYOLOv3算法在保证检测精度的同时,提高了检测速度。展开更多
针对基于You Only Look Once v2算法的目标检测存在精度低及稳健性差的问题,提出一种车辆目标实时检测的You Only Look Once v2优化算法;该算法以You Only Look Once v2算法为基础,通过增加网络深度,增强特征提取能力,同时,通过添加残...针对基于You Only Look Once v2算法的目标检测存在精度低及稳健性差的问题,提出一种车辆目标实时检测的You Only Look Once v2优化算法;该算法以You Only Look Once v2算法为基础,通过增加网络深度,增强特征提取能力,同时,通过添加残差模块,解决网络深度增加带来的梯度消失或弥散问题;该方法将网络结构中低层特征与高层特征进行融合,提升对小目标车辆的检测精度。结果表明,通过在KITTI数据集上进行测试,优化后的算法在检测速度不变的情况下,提高了车辆目标检测精度,平均精度达到0.94,同时提升了小目标检测的准确性。展开更多
The coronavirus(COVID-19)is a lethal virus causing a rapidly infec-tious disease throughout the globe.Spreading awareness,taking preventive mea-sures,imposing strict restrictions on public gatherings,wearing facial ma...The coronavirus(COVID-19)is a lethal virus causing a rapidly infec-tious disease throughout the globe.Spreading awareness,taking preventive mea-sures,imposing strict restrictions on public gatherings,wearing facial masks,and maintaining safe social distancing have become crucial factors in keeping the virus at bay.Even though the world has spent a whole year preventing and curing the disease caused by the COVID-19 virus,the statistics show that the virus can cause an outbreak at any time on a large scale if thorough preventive measures are not maintained accordingly.Tofight the spread of this virus,technologically developed systems have become very useful.However,the implementation of an automatic,robust,continuous,and lightweight monitoring system that can be efficiently deployed on an embedded device still has not become prevalent in the mass community.This paper aims to develop an automatic system to simul-taneously detect social distance and face mask violation in real-time that has been deployed in an embedded system.A modified version of a convolutional neural network,the ResNet50 model,has been utilized to identify masked faces in peo-ple.You Only Look Once(YOLOv3)approach is applied for object detection and the DeepSORT technique is used to measure the social distance.The efficiency of the proposed model is tested on real-time video sequences taken from a video streaming source from an embedded system,Jetson Nano edge computing device,and smartphones,Android and iOS applications.Empirical results show that the implemented model can efficiently detect facial masks and social distance viola-tions with acceptable accuracy and precision scores.展开更多
Radio frequency(RF)-based drone identification technologies have the advantages of long effective distances and low environmental dependence,which has become indispensable for drone surveillance systems.However,since ...Radio frequency(RF)-based drone identification technologies have the advantages of long effective distances and low environmental dependence,which has become indispensable for drone surveillance systems.However,since drones operate in unlicensed frequency bands,a large number of co-frequency devices exist in these bands,which brings a great challenge to traditional signal identification methods.Deep learning techniques provide a new approach to complete endto-end signal identification by directly learning the distribution of RF data.In such scenarios,due to the complexity and high dynamics of the electromagnetic environments,a massive amount of data that can reflect the various propagation conditions of drone signals is necessary for a robust neural network(NN)for identifying drones.In reality,signal acquisition and labeling that meet the above requirements are too costly to implement.Therefore,we propose a virtual electromagnetic environment modeling based data augmentation(DA)method to improve the diversity of drone signal data.The DA method focuses on simulating the spectrograms of drone signals transmitted in real-world environments and randomly generates extra training data in each training epoch.Furthermore,considering the limited processing capability of RF receivers,we modify the original YOLOv5s model to a more lightweight version.Without losing the identification performance,more hardware-friendly designs are applied and the number of parameters decreases about 10-fold.For performance evaluation,we utilized a universal software radio peripheral(USRP)X310 platform to collect RF signals of four drones in an anechoic chamber and a practical wireless scenario.Experiment results reveal that the NN trained with augmented data performs as well as that trained with practical data in the complex electromagnetic environment.展开更多
Aiming at the shortcomings of current gesture tracking methods in accuracy and speed, based on deep learning You Only Look Once version 4(YOLOv4) model, a new YOLOv4 model combined with Kalman filter real-time hand tr...Aiming at the shortcomings of current gesture tracking methods in accuracy and speed, based on deep learning You Only Look Once version 4(YOLOv4) model, a new YOLOv4 model combined with Kalman filter real-time hand tracking method was proposed. The new algorithm can address some problems existing in hand tracking technology such as detection speed, accuracy and stability. The convolutional neural network(CNN) model YOLOv4 is used to detect the target of current frame tracking and Kalman filter is applied to predict the next position and bounding box size of the target according to its current position. The detected target is tracked by comparing the estimated result with the detected target in the next frame and, finally, the real-time hand movement track is displayed. The experimental results validate the proposed algorithm with the overall success rate of 99.43% at speed of 41.822 frame/s, achieving superior results than other algorithms.展开更多
Suspicious fall events are particularly significant hazards for the safety of patients and elders.Recently,suspicious fall event detection has become a robust research case in real-time monitoring.This paper aims to d...Suspicious fall events are particularly significant hazards for the safety of patients and elders.Recently,suspicious fall event detection has become a robust research case in real-time monitoring.This paper aims to detect suspicious fall events during video monitoring of multiple people in different moving back-grounds in an indoor environment;it is further proposed to use a deep learning method known as Long Short Term Memory(LSTM)by introducing visual atten-tion-guided mechanism along with a bi-directional LSTM model.This method contributes essential information on the temporal and spatial locations of‘suspi-cious fall’events in learning the video frame in both forward and backward direc-tions.The effective“You only look once V4”(YOLO V4)–a real-time people detection system illustrates the detection of people in videos,followed by a track-ing module to get their trajectories.Convolutional Neural Network(CNN)fea-tures are extracted for each person tracked through bounding boxes.Subsequently,a visual attention-guided Bi-directional LSTM model is proposed for the final suspicious fall event detection.The proposed method is demonstrated using two different datasets to illustrate the efficiency.The proposed method is evaluated by comparing it with other state-of-the-art methods,showing that it achieves 96.9%accuracy,good performance,and robustness.Hence,it is accep-table to monitor and detect suspicious fall events.展开更多
文摘微动脉瘤是糖尿病视网膜病变的初期症状,消除该病灶可在早期非常有效地预防糖尿病视网膜病变。但由于视网膜结构复杂,同时眼底图像的成像由于患者、环境、采集设备等因素的不同会存在不同的亮度和对比度,现有的微动脉瘤检测算法难以实现该病灶的精确检测和定位,为此本文提出嵌入SENet(squeeze-andexcitation networks)的改进YOLO(you only look once)v4自动检测算法。该算法在YOLOv4网络基础上,首先通过使用一种改进的快速模糊C均值聚类算法对目标样本进行先验框参数优化,以提高先验框与特征图的匹配度;然后,在主干网络嵌入SENet模块,通过强化关键信息,抑制背景信息,提高微动脉瘤的置信度;此外,还在网络颈部增加空间金字塔池化结构以增强主干网络输出特征的接受域,从而有助于分离出重要的上下文信息;最后,在Kaggle数据集上进行模型验证,并与其他方法进行对比。实验结果表明,与其他各种结构的YOLOv4网络模型相比,所提出的嵌入SENet的改进YOLOv4网络模型能显著提高检测结果(与原始YOLOv4相比Fscore提升了12.68%);与其他网络模型以及方法相比,所提出的嵌入SENet的改进YOLOv4网络模型的自动检测精度明显更优,且可实现精准定位。故本文所提出的嵌入SENet的改进YOLOv4算法性能较优,能准确、有效地检测并定位出眼底图像中的微动脉瘤。
文摘针对无人机航拍图像目标检测中视野变化大、时空信息复杂等问题,文中基于YOLOv5(You Only Look Once Version5)架构,提出基于图像低维特征融合的航拍小目标检测模型.引入CA(Coordinate Attention),改进MobileNetV3的反转残差块,增加图像空间维度信息的同时降低模型参数量.改进YOLOv5特征金字塔网络结构,融合浅层网络中的特征图,增加模型对图像低维有效信息的表达能力,进而提升小目标检测精度.同时为了降低航拍图像中复杂背景带来的干扰,引入无参平均注意力模块,同时关注图像的空间注意力与通道注意力;引入VariFocal Loss,降低负样本在训练过程中的权重占比.在VisDrone数据集上的实验验证文中模型的有效性,该模型在有效提升检测精度的同时明显降低复杂度.
基金Tenth Eive-Year Plan of the Committee of the Scientific Tech. of Heilongjiang Province (GB02B4040-1)Heilongjiang Education Bureau (10511012)
文摘In the experiment, the brown rice whose moisture content was 12.5% was used as raw material. The brown rice was grouped, then moisturized differently and milled. While milling, the energy consumption, the rate of broken rice and the crack rate were tested. It is confirmed that the stress crack owing to the moisture added to the brown rice can be avoided when the moisture amount added once is limited to no more than 1.5%. It is also proved that the energy consumption can be reduced, the yielding rate of rice can be increased and that the quality of rice can be improved.
文摘目标检测是计算机视觉领域的研究热点和基础任务,其中基于锚点(Anchor)的目标检测已在众多领域得到广泛应用。当前锚点选取方法主要面临两个问题:基于特定数据集的先验取值尺寸固定、面对不同场景泛化能力弱。计算锚框的无监督K-means算法,受初始值影响较大,对目标尺寸较单一的数据集聚类产生的锚点差异较小,无法充分体现网络多尺度输出的特点。针对上述问题,本文提出一种基于多尺度的目标检测锚点构造方法(multi-scale-anchor,MSA),将聚类产生的锚点根据数据集本身的特性进行尺度的缩放和拉伸,优化的锚点即保留原数据集的特点也体现了模型多尺度的优势。另外,本方法应用在训练的预处理阶段,不增加模型推理时间。最后,选取单阶段主流算法YOLO(You Only Look Once),在多个不同场景的红外或工业场景数据集上进行丰富的实验。结果表明,多尺度锚点优化方法MSA能显著提高小样本场景的检测精度。
文摘针对部署于有限算力平台的YOLOv3(you only look once v3)算法对电容器外观缺陷存在检测速度较慢的问题,提出了基于YOLOv3算法改进的轻量化算法MQYOLOv3。首先采用轻量化网络MobileNet v2作为特征提取模块,通过利用深度可分离式卷积替换一般卷积操作,使得模型的参数量大幅度降低进而提高模型的检测速度,同时也带来了检测精度的降低;然后在网络结构中嵌入空间金字塔池化结构实现局部特征与全局特征的融合、引入距离交并比(distance intersection over union,DIoU)损失函数优化交并比(intersection over union,IoU)损失函数以及使用Mish激活函数优化Leaky ReLU激活函数来提高模型的检测精度。本文采用自制的电容器外观缺陷数据集进行实验,轻量化MQYOLOv3算法的平均精度均值(mean average precision,mAP)为87.96%,较优化前降低了1.16%,检测速度从1.5 FPS提升到7.7 FPS。实验表明,本文设计的轻量化MQYOLOv3算法在保证检测精度的同时,提高了检测速度。
文摘针对基于You Only Look Once v2算法的目标检测存在精度低及稳健性差的问题,提出一种车辆目标实时检测的You Only Look Once v2优化算法;该算法以You Only Look Once v2算法为基础,通过增加网络深度,增强特征提取能力,同时,通过添加残差模块,解决网络深度增加带来的梯度消失或弥散问题;该方法将网络结构中低层特征与高层特征进行融合,提升对小目标车辆的检测精度。结果表明,通过在KITTI数据集上进行测试,优化后的算法在检测速度不变的情况下,提高了车辆目标检测精度,平均精度达到0.94,同时提升了小目标检测的准确性。
文摘The coronavirus(COVID-19)is a lethal virus causing a rapidly infec-tious disease throughout the globe.Spreading awareness,taking preventive mea-sures,imposing strict restrictions on public gatherings,wearing facial masks,and maintaining safe social distancing have become crucial factors in keeping the virus at bay.Even though the world has spent a whole year preventing and curing the disease caused by the COVID-19 virus,the statistics show that the virus can cause an outbreak at any time on a large scale if thorough preventive measures are not maintained accordingly.Tofight the spread of this virus,technologically developed systems have become very useful.However,the implementation of an automatic,robust,continuous,and lightweight monitoring system that can be efficiently deployed on an embedded device still has not become prevalent in the mass community.This paper aims to develop an automatic system to simul-taneously detect social distance and face mask violation in real-time that has been deployed in an embedded system.A modified version of a convolutional neural network,the ResNet50 model,has been utilized to identify masked faces in peo-ple.You Only Look Once(YOLOv3)approach is applied for object detection and the DeepSORT technique is used to measure the social distance.The efficiency of the proposed model is tested on real-time video sequences taken from a video streaming source from an embedded system,Jetson Nano edge computing device,and smartphones,Android and iOS applications.Empirical results show that the implemented model can efficiently detect facial masks and social distance viola-tions with acceptable accuracy and precision scores.
基金supported in part by the Guangzhou Basic and Applied Basic Research Foundation(2023A04J1740)in part by the Shaanxi Provincial Key Research and Development Program(2023-ZDLGY-33,2022ZDLGY05-03,2022ZDLGY05-04)in part by the Fundamental Research Funds for the Central Universities(XJS220116).
文摘Radio frequency(RF)-based drone identification technologies have the advantages of long effective distances and low environmental dependence,which has become indispensable for drone surveillance systems.However,since drones operate in unlicensed frequency bands,a large number of co-frequency devices exist in these bands,which brings a great challenge to traditional signal identification methods.Deep learning techniques provide a new approach to complete endto-end signal identification by directly learning the distribution of RF data.In such scenarios,due to the complexity and high dynamics of the electromagnetic environments,a massive amount of data that can reflect the various propagation conditions of drone signals is necessary for a robust neural network(NN)for identifying drones.In reality,signal acquisition and labeling that meet the above requirements are too costly to implement.Therefore,we propose a virtual electromagnetic environment modeling based data augmentation(DA)method to improve the diversity of drone signal data.The DA method focuses on simulating the spectrograms of drone signals transmitted in real-world environments and randomly generates extra training data in each training epoch.Furthermore,considering the limited processing capability of RF receivers,we modify the original YOLOv5s model to a more lightweight version.Without losing the identification performance,more hardware-friendly designs are applied and the number of parameters decreases about 10-fold.For performance evaluation,we utilized a universal software radio peripheral(USRP)X310 platform to collect RF signals of four drones in an anechoic chamber and a practical wireless scenario.Experiment results reveal that the NN trained with augmented data performs as well as that trained with practical data in the complex electromagnetic environment.
文摘Aiming at the shortcomings of current gesture tracking methods in accuracy and speed, based on deep learning You Only Look Once version 4(YOLOv4) model, a new YOLOv4 model combined with Kalman filter real-time hand tracking method was proposed. The new algorithm can address some problems existing in hand tracking technology such as detection speed, accuracy and stability. The convolutional neural network(CNN) model YOLOv4 is used to detect the target of current frame tracking and Kalman filter is applied to predict the next position and bounding box size of the target according to its current position. The detected target is tracked by comparing the estimated result with the detected target in the next frame and, finally, the real-time hand movement track is displayed. The experimental results validate the proposed algorithm with the overall success rate of 99.43% at speed of 41.822 frame/s, achieving superior results than other algorithms.
文摘Suspicious fall events are particularly significant hazards for the safety of patients and elders.Recently,suspicious fall event detection has become a robust research case in real-time monitoring.This paper aims to detect suspicious fall events during video monitoring of multiple people in different moving back-grounds in an indoor environment;it is further proposed to use a deep learning method known as Long Short Term Memory(LSTM)by introducing visual atten-tion-guided mechanism along with a bi-directional LSTM model.This method contributes essential information on the temporal and spatial locations of‘suspi-cious fall’events in learning the video frame in both forward and backward direc-tions.The effective“You only look once V4”(YOLO V4)–a real-time people detection system illustrates the detection of people in videos,followed by a track-ing module to get their trajectories.Convolutional Neural Network(CNN)fea-tures are extracted for each person tracked through bounding boxes.Subsequently,a visual attention-guided Bi-directional LSTM model is proposed for the final suspicious fall event detection.The proposed method is demonstrated using two different datasets to illustrate the efficiency.The proposed method is evaluated by comparing it with other state-of-the-art methods,showing that it achieves 96.9%accuracy,good performance,and robustness.Hence,it is accep-table to monitor and detect suspicious fall events.
