针对输电线路耐张线夹X射线数字成像(X-ray digital radiography, X-DR)图像检测效率较低,且人工识别易受主观因素影响的问题,提出一种基于YOLO-ISC的输电线路耐张线夹压接缺陷检测方法。首先,在YOLOv8的主干网络中引入注意力特征融合(i...针对输电线路耐张线夹X射线数字成像(X-ray digital radiography, X-DR)图像检测效率较低,且人工识别易受主观因素影响的问题,提出一种基于YOLO-ISC的输电线路耐张线夹压接缺陷检测方法。首先,在YOLOv8的主干网络中引入注意力特征融合(iterative attention feature fusion, iAFF)模块,通过逐层融合不同尺度的特征减少信息的丢失;其次,采用SimAM注意力机制、内容感知特征重组算子(content-aware reassembly of features, CARAFE)构建融合网络PANet-SC,增强了缺陷特征的表达能力;最后,将融合后的特征输入检测头YOLO Head进行分类预测。实验结果表明,所用YOLO-ISC模型在检测耐张线夹X-DR图像的平均检测精度(mean average precision, mAP)值达到92.49%,检测速度为23帧/s。针对某类缺陷检测精度不足的问题,讨论模型置信度阈值对实际检测结果的影响,降低模型的误检率。将检测结果与SSD、Faster RCNN、DETR、YOLOv8等算法进行比较,验证所用方法的有效性。展开更多
With the rise in drowning accidents in swimming pools,the demand for the precision and speed in artificial intelligence(AI)drowning detection methods has become increasingly crucial.Here,an improved YOLO-based method,...With the rise in drowning accidents in swimming pools,the demand for the precision and speed in artificial intelligence(AI)drowning detection methods has become increasingly crucial.Here,an improved YOLO-based method,named DrownACB-YOLO,for drowning detection in swimming pools is proposed.Since existing methods focus on the drowned state,a transition label is added to the original dataset to provide timely alerts.Following this expanded dataset,two improvements are implemented in the original YOLOv5.Firstly,the spatial pyramid pooling(SPP)module and the default upsampling operator are replaced by the atrous spatial pyramid pooling(ASPP)module and the content-aware reassembly of feature(CARAFE)module,respectively.Secondly,the cross stage partial bottleneck with three convolutions(C3)module at the end of the backbone is replaced with the bottleneck transformer(BotNet)module.The results of comparison experiments demonstrate that DrownACB-YOLO performs better than other models.展开更多
【目的】为了解决红外热成像缺陷检测中缺陷信号微弱和边界模糊导致的检测准确性与鲁棒性不足的问题,提出一种基于增强图像引导的改进YOLOv11检测模型(improved YOLOv11 detection model based on enhanced image guidance,IEIG-YOLOv11...【目的】为了解决红外热成像缺陷检测中缺陷信号微弱和边界模糊导致的检测准确性与鲁棒性不足的问题,提出一种基于增强图像引导的改进YOLOv11检测模型(improved YOLOv11 detection model based on enhanced image guidance,IEIG-YOLOv11)。【方法】首先,在原始YOLOv11模型的输入端引入预处理后的增强热图像,丰富输入特征表达;然后,在主干网络引入高效多尺度注意力机制(efficient multi-scale attention,EMA),结合残差连接提高浅层特征表达能力,增强对微弱缺陷的检测能力;最后,在颈部网络中引入内容感知特征重组(content-aware reassembly of features,CARAFE)上采样模块,以提高浅层特征重建质量,强化缺陷细节特征。【结果】经消融试验验证,上述改进均对模型性能有所提升。针对缺陷目标检测,IEIG-YOLOv11模型的精确率、召回率及F 1分数分别达到0.967、0.978和0.972。IEIG-YOLOv11模型与主流目标检测模型的对比试验结果表明,IEIG-YOLOv11模型的各项性能均优于其他模型,显著减少了漏检现象,有效增强了对边缘模糊及微弱缺陷的识别能力,并能兼顾检测精度与效率。【结论】IEIG-YOLOv11模型为实现微弱信号情况下的红外缺陷目标检测提供了解决方案,进而为红外热成像技术的实际应用奠定了基础。展开更多
在自动驾驶场景中,针对复杂背景对车辆和行人检测目标影响大、小目标检测精度不高的问题,提出一种基于内容感知重组特征和自适应融合的YOLOv5(content-aware reassembly of feature and adaptive fusion YOLOv5,CRAF-YOLOv5)车辆及行人...在自动驾驶场景中,针对复杂背景对车辆和行人检测目标影响大、小目标检测精度不高的问题,提出一种基于内容感知重组特征和自适应融合的YOLOv5(content-aware reassembly of feature and adaptive fusion YOLOv5,CRAF-YOLOv5)车辆及行人检测算法。通过引入通道注意力机制形成多通道特征提取网络,增强复杂背景下目标特征的提取性能;在特征融合前段,通过内容感知重组特征进行上采样,并添加基于跳跃连接结构,强化浅层网络对小目标特征的表征能力;在特征融合后段,采用自适应权重融合方式学习不同尺度特征,实现深层和浅层特征的动态学习和深度融合。实验结果表明,该算法在BDD100K和KITTI数据集上车辆行人目标检测平均均值精度分别达到84.40%和93.35%,较YOLOv5基准算法分别提高了3.90%和0.45%。展开更多
文摘With the rise in drowning accidents in swimming pools,the demand for the precision and speed in artificial intelligence(AI)drowning detection methods has become increasingly crucial.Here,an improved YOLO-based method,named DrownACB-YOLO,for drowning detection in swimming pools is proposed.Since existing methods focus on the drowned state,a transition label is added to the original dataset to provide timely alerts.Following this expanded dataset,two improvements are implemented in the original YOLOv5.Firstly,the spatial pyramid pooling(SPP)module and the default upsampling operator are replaced by the atrous spatial pyramid pooling(ASPP)module and the content-aware reassembly of feature(CARAFE)module,respectively.Secondly,the cross stage partial bottleneck with three convolutions(C3)module at the end of the backbone is replaced with the bottleneck transformer(BotNet)module.The results of comparison experiments demonstrate that DrownACB-YOLO performs better than other models.
文摘【目的】为了解决红外热成像缺陷检测中缺陷信号微弱和边界模糊导致的检测准确性与鲁棒性不足的问题,提出一种基于增强图像引导的改进YOLOv11检测模型(improved YOLOv11 detection model based on enhanced image guidance,IEIG-YOLOv11)。【方法】首先,在原始YOLOv11模型的输入端引入预处理后的增强热图像,丰富输入特征表达;然后,在主干网络引入高效多尺度注意力机制(efficient multi-scale attention,EMA),结合残差连接提高浅层特征表达能力,增强对微弱缺陷的检测能力;最后,在颈部网络中引入内容感知特征重组(content-aware reassembly of features,CARAFE)上采样模块,以提高浅层特征重建质量,强化缺陷细节特征。【结果】经消融试验验证,上述改进均对模型性能有所提升。针对缺陷目标检测,IEIG-YOLOv11模型的精确率、召回率及F 1分数分别达到0.967、0.978和0.972。IEIG-YOLOv11模型与主流目标检测模型的对比试验结果表明,IEIG-YOLOv11模型的各项性能均优于其他模型,显著减少了漏检现象,有效增强了对边缘模糊及微弱缺陷的识别能力,并能兼顾检测精度与效率。【结论】IEIG-YOLOv11模型为实现微弱信号情况下的红外缺陷目标检测提供了解决方案,进而为红外热成像技术的实际应用奠定了基础。
文摘在自动驾驶场景中,针对复杂背景对车辆和行人检测目标影响大、小目标检测精度不高的问题,提出一种基于内容感知重组特征和自适应融合的YOLOv5(content-aware reassembly of feature and adaptive fusion YOLOv5,CRAF-YOLOv5)车辆及行人检测算法。通过引入通道注意力机制形成多通道特征提取网络,增强复杂背景下目标特征的提取性能;在特征融合前段,通过内容感知重组特征进行上采样,并添加基于跳跃连接结构,强化浅层网络对小目标特征的表征能力;在特征融合后段,采用自适应权重融合方式学习不同尺度特征,实现深层和浅层特征的动态学习和深度融合。实验结果表明,该算法在BDD100K和KITTI数据集上车辆行人目标检测平均均值精度分别达到84.40%和93.35%,较YOLOv5基准算法分别提高了3.90%和0.45%。
