现有的注意力机制仅增强特征图的通道或空间维度,未能充分捕捉细微视觉元素和多尺度特征变化。为解决此问题,提出一种基于局部分块与全局多尺度特征融合的注意力机制(patch and global multiscale attention,PGMA)。将特征图分割成多个...现有的注意力机制仅增强特征图的通道或空间维度,未能充分捕捉细微视觉元素和多尺度特征变化。为解决此问题,提出一种基于局部分块与全局多尺度特征融合的注意力机制(patch and global multiscale attention,PGMA)。将特征图分割成多个小块,分别计算这些小块的注意力得分,增强对局部信息的感知能力。使用一组空洞卷积计算整个特征图的得分,获得全局多尺度信息的权衡。实验中,将PGMA集成到U-Net、DeepLab、SegNet等语义分割网络中,有效提升了它们的分割性能。这表明PGMA在增强CNN性能方面优于当前主流方法。展开更多
随着遥感图像中小目标检测问题的日益突出,传统目标检测方法在小目标的精确定位上存在局限性。为解决这一问题,本文提出了一种基于YOLOv11模型的多尺度注意力机制优化方法。首先,删除了YOLOv11模型中用于大目标检测的20 × 20尺度...随着遥感图像中小目标检测问题的日益突出,传统目标检测方法在小目标的精确定位上存在局限性。为解决这一问题,本文提出了一种基于YOLOv11模型的多尺度注意力机制优化方法。首先,删除了YOLOv11模型中用于大目标检测的20 × 20尺度检测层,增加了160 × 160尺度的小目标检测层,以提升小目标的检测精度。其次,采用EIoU (Enhanced Intersection over Union)损失函数替代CIoU损失函数,解决了CIoU在长宽比差异较大的目标中的定位问题,从而加速收敛并提高定位精度。最后,结合空间注意力和通道注意力机制,增强了模型对不同尺度目标的感知能力。实验结果表明,优化后的YOLOv11模型在多个遥感图像数据集上表现出较传统YOLOv11显著提高的精度、召回率和F1分数,特别在小目标检测任务中具有更强的鲁棒性和更高的检测精度。研究表明,提出的方法能有效提升小目标检测性能,为遥感图像分析提供了新的解决方案。With the increasingly prominent problem of small target detection in remote sensing images, traditional object detection methods have limitations in accurately locating small targets. To address this issue, this paper proposes a multi-scale attention mechanism optimization method based on the YOLOv11 model. Firstly, the 20 × 20 scale detection layer used for large object detection in the YOLOv11 model was removed, and a 160 × 160 scale small object detection layer was added to improve the detection accuracy of small objects. Secondly, the EIoU (Enhanced Intersection over Union) loss function is used instead of the CIoU loss function to solve the localization problem of CIoU in targets with large aspect ratio differences, thereby accelerating convergence and improving localization accuracy. Finally, by combining spatial attention and channel attention mechanisms, the model’s perception ability for targets of different scales was enhanced. The experimental results show that the optimized YOLOv11 model exhibits significantly improved accuracy, recall, and F1 score compared to traditional YOLOv11 on multiple remote sensing image datasets, especially in small object detection tasks with stronger robustness and higher detection accuracy. Research has shown that the proposed method can effectively improve the performance of small object detection, providing a new solution for remote sensing image analysis.展开更多
【目的】设计MobileNet with large convolution Unit(Mobile-LU)模型,解决由于辣椒病害种类复杂和类间差异不明显而造成的病害识别困难、准确率低等问题。【方法】重新构建MobileNetV3的特征提取层,在并行分支单元结构中采用不同尺度...【目的】设计MobileNet with large convolution Unit(Mobile-LU)模型,解决由于辣椒病害种类复杂和类间差异不明显而造成的病害识别困难、准确率低等问题。【方法】重新构建MobileNetV3的特征提取层,在并行分支单元结构中采用不同尺度的分离卷积,增强模型对辣椒病害尺寸差异特征的表达能力;引入Squeeze-and-Excitation(SE)注意力机制,加强模型对病害相关的特征的学习,提高病害识别准确率;同时使用Leaky ReLU激活函数,在负值区域引入小的斜率,避免网络神经元死亡问题;调整输出层节点个数,更好适应辣椒病害分类任务。【结果】Mobile-LU模型的识别准确率达到98.2%,相较于MobilenetV3-small、ResNet34、VGG16、Alexnet、Swin Transformer、MobileVIT等模型分别高出8.9、7.3、4.4、20.4、6.0、8.3个百分点,且Mobile-LU模型在精确率、召回率、特异度以及F1分数等关键性能指标上也均有优势。【结论】Mobile-LU模型对辣椒病害的识别性能更优,能更好满足辣椒病害识别任务。展开更多
文摘现有的注意力机制仅增强特征图的通道或空间维度,未能充分捕捉细微视觉元素和多尺度特征变化。为解决此问题,提出一种基于局部分块与全局多尺度特征融合的注意力机制(patch and global multiscale attention,PGMA)。将特征图分割成多个小块,分别计算这些小块的注意力得分,增强对局部信息的感知能力。使用一组空洞卷积计算整个特征图的得分,获得全局多尺度信息的权衡。实验中,将PGMA集成到U-Net、DeepLab、SegNet等语义分割网络中,有效提升了它们的分割性能。这表明PGMA在增强CNN性能方面优于当前主流方法。
文摘随着遥感图像中小目标检测问题的日益突出,传统目标检测方法在小目标的精确定位上存在局限性。为解决这一问题,本文提出了一种基于YOLOv11模型的多尺度注意力机制优化方法。首先,删除了YOLOv11模型中用于大目标检测的20 × 20尺度检测层,增加了160 × 160尺度的小目标检测层,以提升小目标的检测精度。其次,采用EIoU (Enhanced Intersection over Union)损失函数替代CIoU损失函数,解决了CIoU在长宽比差异较大的目标中的定位问题,从而加速收敛并提高定位精度。最后,结合空间注意力和通道注意力机制,增强了模型对不同尺度目标的感知能力。实验结果表明,优化后的YOLOv11模型在多个遥感图像数据集上表现出较传统YOLOv11显著提高的精度、召回率和F1分数,特别在小目标检测任务中具有更强的鲁棒性和更高的检测精度。研究表明,提出的方法能有效提升小目标检测性能,为遥感图像分析提供了新的解决方案。With the increasingly prominent problem of small target detection in remote sensing images, traditional object detection methods have limitations in accurately locating small targets. To address this issue, this paper proposes a multi-scale attention mechanism optimization method based on the YOLOv11 model. Firstly, the 20 × 20 scale detection layer used for large object detection in the YOLOv11 model was removed, and a 160 × 160 scale small object detection layer was added to improve the detection accuracy of small objects. Secondly, the EIoU (Enhanced Intersection over Union) loss function is used instead of the CIoU loss function to solve the localization problem of CIoU in targets with large aspect ratio differences, thereby accelerating convergence and improving localization accuracy. Finally, by combining spatial attention and channel attention mechanisms, the model’s perception ability for targets of different scales was enhanced. The experimental results show that the optimized YOLOv11 model exhibits significantly improved accuracy, recall, and F1 score compared to traditional YOLOv11 on multiple remote sensing image datasets, especially in small object detection tasks with stronger robustness and higher detection accuracy. Research has shown that the proposed method can effectively improve the performance of small object detection, providing a new solution for remote sensing image analysis.
文摘【目的】设计MobileNet with large convolution Unit(Mobile-LU)模型,解决由于辣椒病害种类复杂和类间差异不明显而造成的病害识别困难、准确率低等问题。【方法】重新构建MobileNetV3的特征提取层,在并行分支单元结构中采用不同尺度的分离卷积,增强模型对辣椒病害尺寸差异特征的表达能力;引入Squeeze-and-Excitation(SE)注意力机制,加强模型对病害相关的特征的学习,提高病害识别准确率;同时使用Leaky ReLU激活函数,在负值区域引入小的斜率,避免网络神经元死亡问题;调整输出层节点个数,更好适应辣椒病害分类任务。【结果】Mobile-LU模型的识别准确率达到98.2%,相较于MobilenetV3-small、ResNet34、VGG16、Alexnet、Swin Transformer、MobileVIT等模型分别高出8.9、7.3、4.4、20.4、6.0、8.3个百分点,且Mobile-LU模型在精确率、召回率、特异度以及F1分数等关键性能指标上也均有优势。【结论】Mobile-LU模型对辣椒病害的识别性能更优,能更好满足辣椒病害识别任务。
