机器阅读理解(Machine reading comprehension,MRC)是自然语言处理领域中一项重要研究任务,其目标是通过机器理解给定的阅读材料和问题,最终实现自动答题.目前联合观点类问题解答和答案依据挖掘的多任务联合学习研究在机器阅读理解应用...机器阅读理解(Machine reading comprehension,MRC)是自然语言处理领域中一项重要研究任务,其目标是通过机器理解给定的阅读材料和问题,最终实现自动答题.目前联合观点类问题解答和答案依据挖掘的多任务联合学习研究在机器阅读理解应用中受到广泛关注,它可以同时给出问题答案和支撑答案的相关证据,然而现有观点类问题的答题方法在答案线索识别上表现还不是太好,已有答案依据挖掘方法仍不能较好捕获段落中词语之间的依存关系.基于此,引入多头自注意力(Multi-head self-attention,MHSA)进一步挖掘阅读材料中观点类问题的文字线索,改进了观点类问题的自动解答方法;将句法关系融入到图构建过程中,提出了基于关联要素关系图的多跳推理方法,实现了答案支撑句挖掘;通过联合优化两个子任务,构建了基于多任务联合学习的阅读理解模型.在2020中国“法研杯”司法人工智能挑战赛(China AI Law Challenge 2020,CAIL2020)和HotpotQA数据集上的实验结果表明,本文提出的方法比已有基线模型的效果更好.展开更多
目的桨叶运动参数是直升机设计到生产的重要指标,传统的视觉测量方法直接应用于室外环境下,由于受复杂光照背景影响,存在找不到桨叶区域、不能进行准确测量的问题。据此,本文提出一种融合多特征与自注意力的旋转目标检测器(fusion multi...目的桨叶运动参数是直升机设计到生产的重要指标,传统的视觉测量方法直接应用于室外环境下,由于受复杂光照背景影响,存在找不到桨叶区域、不能进行准确测量的问题。据此,本文提出一种融合多特征与自注意力的旋转目标检测器(fusion multi-feature and self-attention rotating detector,FMSA-RD)。方法首先,针对YOLOv5s(you only look once)特征提取能力不足和冗余问题,在主干网络中设计了更为有效的多特征提取和融合模块,结合不同时刻位置与尺度下的特征信息以提高网络对室外桨叶的检测精度;并去掉部分无关卷积层以简化模块结构参数。其次,融合多头自注意力机制与CSP(crossstage partial convolution)瓶颈结构,整合全局信息以抑制室外复杂光照背景干扰。最后,引入倾斜交并比(skew intersection over union,SKEWIOU)损失和角度损失,改进损失函数,进一步提升桨叶检测精度。结果本文进行了多组对比实验,分别在自制的室外直升机桨叶数据集和公共数据集DOTA-v1.0(dataset for object detection in aerial images)上进行验证,对比基线YOLOv5s目标检测网络,本文模型平均精度均值(mean average precision,mAP)分别提高6.6%和12.8%,帧速率(frames per second,FPS)分别提高21.8%和47.7%。结论本文设计的旋转目标检测模型,提升了室外复杂光照背景下桨叶的检测精度和速度。展开更多
Cloud detection is a critical preprocessing step in remote sensing image processing, as the presence of clouds significantly affects the accuracy of remote sensing data and limits its applicability across various doma...Cloud detection is a critical preprocessing step in remote sensing image processing, as the presence of clouds significantly affects the accuracy of remote sensing data and limits its applicability across various domains. This study presents an enhanced cloud detection method based on the U-Net architecture, designed to address the challenges of multi-scale cloud features and long-range dependencies inherent in remote sensing imagery. A Multi-Scale Dilated Attention (MSDA) module is introduced to effectively integrate multi-scale information and model long-range dependencies across different scales, enhancing the model’s ability to detect clouds of varying sizes. Additionally, a Multi-Head Self-Attention (MHSA) mechanism is incorporated to improve the model’s capacity for capturing finer details, particularly in distinguishing thin clouds from surface features. A multi-path supervision mechanism is also devised to ensure the model learns cloud features at multiple scales, further boosting the accuracy and robustness of cloud mask generation. Experimental results demonstrate that the enhanced model achieves superior performance compared to other benchmarked methods in complex scenarios. It significantly improves cloud detection accuracy, highlighting its strong potential for practical applications in cloud detection tasks.展开更多
Vehicle detection is a crucial aspect of intelligent transportation systems(ITS)and autonomous driving technologies.The complexity and diversity of real-world road environments,coupled with traffic congestion,pose sig...Vehicle detection is a crucial aspect of intelligent transportation systems(ITS)and autonomous driving technologies.The complexity and diversity of real-world road environments,coupled with traffic congestion,pose significant challenges to the accuracy and real-time performance of vehicle detection models.To address these challenges,this paper introduces a fast and accurate vehicle detection algorithm named BES-Net.Firstly,the BoTNet module is integrated into the backbone network to bolster the model’s long-distance dependency,address the complexities and diversity of road environments,and accelerate the detection speed of the BES-Net network.Secondly,to accommodate the varying sizes of target vehicles,the efficient multi-scale attention mechanism(EMA)was added to enrich feature information and enhance the model’s expressive power by combining features from multiple scales.Finally,the Slide loss function is specifically designed to give higher weight to difficult samples,thereby improving the detection accuracy of small targets.The experimental results demonstrate that BES-Net achieves superior performance compared to conventional object detection models,with mAP50(mean Average Precision),precision,and recall reaching 73.2%,74.8%,and 73.1%,respectively.These metrics represent significant improvements of 8.5%,7.2%,and 11.7%over the baseline network.This significant improvement underscores the high robustness of the BES-Net model in vehicle detection tasks.In addition,the BES-Net network has been deployed on NVIDIA Jetson Orin NX equipment,providing a solid foundation for its integration into intelligent transportation systems.This deployment not only showcases the practicality of the model but also highlights its potential for real-world applications in autonomous driving and ITS.展开更多
文摘机器阅读理解(Machine reading comprehension,MRC)是自然语言处理领域中一项重要研究任务,其目标是通过机器理解给定的阅读材料和问题,最终实现自动答题.目前联合观点类问题解答和答案依据挖掘的多任务联合学习研究在机器阅读理解应用中受到广泛关注,它可以同时给出问题答案和支撑答案的相关证据,然而现有观点类问题的答题方法在答案线索识别上表现还不是太好,已有答案依据挖掘方法仍不能较好捕获段落中词语之间的依存关系.基于此,引入多头自注意力(Multi-head self-attention,MHSA)进一步挖掘阅读材料中观点类问题的文字线索,改进了观点类问题的自动解答方法;将句法关系融入到图构建过程中,提出了基于关联要素关系图的多跳推理方法,实现了答案支撑句挖掘;通过联合优化两个子任务,构建了基于多任务联合学习的阅读理解模型.在2020中国“法研杯”司法人工智能挑战赛(China AI Law Challenge 2020,CAIL2020)和HotpotQA数据集上的实验结果表明,本文提出的方法比已有基线模型的效果更好.
文摘目的桨叶运动参数是直升机设计到生产的重要指标,传统的视觉测量方法直接应用于室外环境下,由于受复杂光照背景影响,存在找不到桨叶区域、不能进行准确测量的问题。据此,本文提出一种融合多特征与自注意力的旋转目标检测器(fusion multi-feature and self-attention rotating detector,FMSA-RD)。方法首先,针对YOLOv5s(you only look once)特征提取能力不足和冗余问题,在主干网络中设计了更为有效的多特征提取和融合模块,结合不同时刻位置与尺度下的特征信息以提高网络对室外桨叶的检测精度;并去掉部分无关卷积层以简化模块结构参数。其次,融合多头自注意力机制与CSP(crossstage partial convolution)瓶颈结构,整合全局信息以抑制室外复杂光照背景干扰。最后,引入倾斜交并比(skew intersection over union,SKEWIOU)损失和角度损失,改进损失函数,进一步提升桨叶检测精度。结果本文进行了多组对比实验,分别在自制的室外直升机桨叶数据集和公共数据集DOTA-v1.0(dataset for object detection in aerial images)上进行验证,对比基线YOLOv5s目标检测网络,本文模型平均精度均值(mean average precision,mAP)分别提高6.6%和12.8%,帧速率(frames per second,FPS)分别提高21.8%和47.7%。结论本文设计的旋转目标检测模型,提升了室外复杂光照背景下桨叶的检测精度和速度。
文摘Cloud detection is a critical preprocessing step in remote sensing image processing, as the presence of clouds significantly affects the accuracy of remote sensing data and limits its applicability across various domains. This study presents an enhanced cloud detection method based on the U-Net architecture, designed to address the challenges of multi-scale cloud features and long-range dependencies inherent in remote sensing imagery. A Multi-Scale Dilated Attention (MSDA) module is introduced to effectively integrate multi-scale information and model long-range dependencies across different scales, enhancing the model’s ability to detect clouds of varying sizes. Additionally, a Multi-Head Self-Attention (MHSA) mechanism is incorporated to improve the model’s capacity for capturing finer details, particularly in distinguishing thin clouds from surface features. A multi-path supervision mechanism is also devised to ensure the model learns cloud features at multiple scales, further boosting the accuracy and robustness of cloud mask generation. Experimental results demonstrate that the enhanced model achieves superior performance compared to other benchmarked methods in complex scenarios. It significantly improves cloud detection accuracy, highlighting its strong potential for practical applications in cloud detection tasks.
基金funded by National Natural Science Foundation of China(No.61961011)Innovation Project of Guangxi Graduate Education No.YCSW2025411.
文摘Vehicle detection is a crucial aspect of intelligent transportation systems(ITS)and autonomous driving technologies.The complexity and diversity of real-world road environments,coupled with traffic congestion,pose significant challenges to the accuracy and real-time performance of vehicle detection models.To address these challenges,this paper introduces a fast and accurate vehicle detection algorithm named BES-Net.Firstly,the BoTNet module is integrated into the backbone network to bolster the model’s long-distance dependency,address the complexities and diversity of road environments,and accelerate the detection speed of the BES-Net network.Secondly,to accommodate the varying sizes of target vehicles,the efficient multi-scale attention mechanism(EMA)was added to enrich feature information and enhance the model’s expressive power by combining features from multiple scales.Finally,the Slide loss function is specifically designed to give higher weight to difficult samples,thereby improving the detection accuracy of small targets.The experimental results demonstrate that BES-Net achieves superior performance compared to conventional object detection models,with mAP50(mean Average Precision),precision,and recall reaching 73.2%,74.8%,and 73.1%,respectively.These metrics represent significant improvements of 8.5%,7.2%,and 11.7%over the baseline network.This significant improvement underscores the high robustness of the BES-Net model in vehicle detection tasks.In addition,the BES-Net network has been deployed on NVIDIA Jetson Orin NX equipment,providing a solid foundation for its integration into intelligent transportation systems.This deployment not only showcases the practicality of the model but also highlights its potential for real-world applications in autonomous driving and ITS.
