Underwater imaging is frequently influenced by factors such as illumination,scattering,and refraction,which can result in low image contrast and blurriness.Moreover,the presence of numerous small,overlapping targets r...Underwater imaging is frequently influenced by factors such as illumination,scattering,and refraction,which can result in low image contrast and blurriness.Moreover,the presence of numerous small,overlapping targets reduces detection accuracy.To address these challenges,first,green channel images are preprocessed to rectify color bias while improving contrast and clarity.Se-cond,the YOLO-DBS network that employs deformable convolution is proposed to enhance feature learning from underwater blurry images.The ECA attention mechanism is also introduced to strengthen feature focus.Moreover,a bidirectional feature pyramid net-work is utilized for efficient multilayer feature fusion while removing nodes that contribute minimally to detection performance.In addition,the SIoU loss function that considers factors such as angular error and distance deviation is incorporated into the network.Validation on the RUOD dataset demonstrates that YOLO-DBS achieves approximately 3.1%improvement in mAP@0.5 compared with YOLOv8n and surpasses YOLOv9-tiny by 1.3%.YOLO-DBS reduces parameter count by 32%relative to YOLOv8n,thereby demonstrating superior performance in real-time detection on underwater observation platforms.展开更多
针对印刷电路板(Printed Circuit Board,PCB)表面缺陷检测任务中模型体积和参数量较大的问题,提出了一种基于通道剪枝的轻量级YOLOv8n网络PCB缺陷检测算法。为有效提升对PCB小目标缺陷的特征提取能力,采用RepViT作为特征提取网络;为提...针对印刷电路板(Printed Circuit Board,PCB)表面缺陷检测任务中模型体积和参数量较大的问题,提出了一种基于通道剪枝的轻量级YOLOv8n网络PCB缺陷检测算法。为有效提升对PCB小目标缺陷的特征提取能力,采用RepViT作为特征提取网络;为提升网络对小目标的关注度,减少神经网络推理过程中的梯度信息重复,将颈部网络的卷积模块替换为Rep-Net with Cross-Stage Partial CSP and ELAN(RepNCSPELAN4);为降低缺陷重叠时检测框失真现象,在预测部分使用Focaler-MPDIoU替换完全交并比(Complete Intersection over Union,CIoU);利用层自适应幅度分数剪枝(Layer Adaptive Magnitude based Pruning,LAMP)方法对融合改进方法的模型进行修剪,去除模型中冗余的梯度信息和权重,减少参数量和浮点运算量,压缩模型体积。实验结果表明,在PCB公开数据集中,经过LAMP之后,该算法相较于YOLOv8n,参数量下降60.8%,模型体积减小50.8%,计算量下降48.8%,平均精度均值(mean Average Precision,mAP)提高3.8%。在提高精度的同时,计算量、参数量和模型体积都低于原模型,满足在低配置设备下的使用需求。展开更多
由于合成孔径雷达独特的成像机制,现有的深度学习检测算法,难以在精度和速度之间找到良好的平衡。因此针对边缘端应用需求,结合剪枝方法设计了一种更轻量的SAR图像飞机目标检测网络SAERFDnet。其以YOLOv8n模型为基线,主干使用重参数化...由于合成孔径雷达独特的成像机制,现有的深度学习检测算法,难以在精度和速度之间找到良好的平衡。因此针对边缘端应用需求,结合剪枝方法设计了一种更轻量的SAR图像飞机目标检测网络SAERFDnet。其以YOLOv8n模型为基线,主干使用重参数化大核卷积进行特征提取,颈部使用自适应多尺度离散特征融合模块,以较浅的网络深度获得更大的有效感受野。其次,改进网络在检测头分类分支引入可变形卷积,使得网络更关注不同类别目标的几何特征差异,在回归分支引入频率自适应扩张卷积,加强对图像高频区域的目标定位能力。最后,使用了模型剪枝技术获得了更轻量高效的模型。采用3个公开的实测数据集进行实验,在SAR-AIRcraft-1.0数据集上的结果表明,该方法以0.5 M参数量和2 G FLOPS的参数量和计算量达到了96.3%mAP50和72.5%mAP50-95,相比YOLOv8n模型参数量降低83.3%,计算量降低75.3%,同时提高了0.7%mAP50和2.2%mAP50-95的检测精度,对比其他模型检测结果,该方法能在保证检测精度的条件下,有效提升SAR图像飞机目标检测的检测效率。此外,在SADD数据集和高分三号飞机目标数据集上进行了迁移实验,结果均表明该方法具有良好的泛化性和鲁棒性。展开更多
基金funded by the Jilin City Science and Technology Innovation Development Plan Project(No.20240302014)the Jilin Provincial Department of Educa-tion Science and Technology Research Project(No.JJKH 20250879KJ)the Jilin Province Science and Tech-nology Development Plan Project(No.YDZJ202401640 ZYTS).
文摘Underwater imaging is frequently influenced by factors such as illumination,scattering,and refraction,which can result in low image contrast and blurriness.Moreover,the presence of numerous small,overlapping targets reduces detection accuracy.To address these challenges,first,green channel images are preprocessed to rectify color bias while improving contrast and clarity.Se-cond,the YOLO-DBS network that employs deformable convolution is proposed to enhance feature learning from underwater blurry images.The ECA attention mechanism is also introduced to strengthen feature focus.Moreover,a bidirectional feature pyramid net-work is utilized for efficient multilayer feature fusion while removing nodes that contribute minimally to detection performance.In addition,the SIoU loss function that considers factors such as angular error and distance deviation is incorporated into the network.Validation on the RUOD dataset demonstrates that YOLO-DBS achieves approximately 3.1%improvement in mAP@0.5 compared with YOLOv8n and surpasses YOLOv9-tiny by 1.3%.YOLO-DBS reduces parameter count by 32%relative to YOLOv8n,thereby demonstrating superior performance in real-time detection on underwater observation platforms.
文摘针对印刷电路板(Printed Circuit Board,PCB)表面缺陷检测任务中模型体积和参数量较大的问题,提出了一种基于通道剪枝的轻量级YOLOv8n网络PCB缺陷检测算法。为有效提升对PCB小目标缺陷的特征提取能力,采用RepViT作为特征提取网络;为提升网络对小目标的关注度,减少神经网络推理过程中的梯度信息重复,将颈部网络的卷积模块替换为Rep-Net with Cross-Stage Partial CSP and ELAN(RepNCSPELAN4);为降低缺陷重叠时检测框失真现象,在预测部分使用Focaler-MPDIoU替换完全交并比(Complete Intersection over Union,CIoU);利用层自适应幅度分数剪枝(Layer Adaptive Magnitude based Pruning,LAMP)方法对融合改进方法的模型进行修剪,去除模型中冗余的梯度信息和权重,减少参数量和浮点运算量,压缩模型体积。实验结果表明,在PCB公开数据集中,经过LAMP之后,该算法相较于YOLOv8n,参数量下降60.8%,模型体积减小50.8%,计算量下降48.8%,平均精度均值(mean Average Precision,mAP)提高3.8%。在提高精度的同时,计算量、参数量和模型体积都低于原模型,满足在低配置设备下的使用需求。
文摘由于合成孔径雷达独特的成像机制,现有的深度学习检测算法,难以在精度和速度之间找到良好的平衡。因此针对边缘端应用需求,结合剪枝方法设计了一种更轻量的SAR图像飞机目标检测网络SAERFDnet。其以YOLOv8n模型为基线,主干使用重参数化大核卷积进行特征提取,颈部使用自适应多尺度离散特征融合模块,以较浅的网络深度获得更大的有效感受野。其次,改进网络在检测头分类分支引入可变形卷积,使得网络更关注不同类别目标的几何特征差异,在回归分支引入频率自适应扩张卷积,加强对图像高频区域的目标定位能力。最后,使用了模型剪枝技术获得了更轻量高效的模型。采用3个公开的实测数据集进行实验,在SAR-AIRcraft-1.0数据集上的结果表明,该方法以0.5 M参数量和2 G FLOPS的参数量和计算量达到了96.3%mAP50和72.5%mAP50-95,相比YOLOv8n模型参数量降低83.3%,计算量降低75.3%,同时提高了0.7%mAP50和2.2%mAP50-95的检测精度,对比其他模型检测结果,该方法能在保证检测精度的条件下,有效提升SAR图像飞机目标检测的检测效率。此外,在SADD数据集和高分三号飞机目标数据集上进行了迁移实验,结果均表明该方法具有良好的泛化性和鲁棒性。
