针对光照不均匀和水表表盘雾化的指针式水表在读数检测时出现漏检、误检等问题,提出一种基于改进YOLOv5s的指针式水表读数检测方法。首先,采用Mosaic、Mixup等数据增强方法,提高模型的泛化能力;其次,引入加权双向特征金字塔网络(bilater...针对光照不均匀和水表表盘雾化的指针式水表在读数检测时出现漏检、误检等问题,提出一种基于改进YOLOv5s的指针式水表读数检测方法。首先,采用Mosaic、Mixup等数据增强方法,提高模型的泛化能力;其次,引入加权双向特征金字塔网络(bilateral feature pyramid network, BiFPN)实现更高层次的特征融合使得水表图像的深层特征图和浅层特征图充分融合,提高网络的表达能力;然后,嵌入卷积注意力机制(convolutional block attention module, CBAM),在通道和空间双重维度上强化指针式水表子表盘示数特征;最后将完全交并比损失函数(complete intersection over union loss, CIoU-Loss)替换为SIoU_Loss(scylla intersection over union loss),提升边界框的回归精度。改进算法的mAP@0.5达到97.8%,比YOLOv5s原始网络提升了3.2%。实验结果表明:该算法能有效提高指针式水表的读数检测精度。展开更多
针对现有仪表读数方法易受光照不均等因素影响,而导致读数误差大的问题,提出一种基于深度学习的全自动指针式仪表读数方法。首先,引入YOLOv7网络提取表盘区域;其次,采用文中提出的VCA-UNet(VGG16Net,improved skip connections and ASPP...针对现有仪表读数方法易受光照不均等因素影响,而导致读数误差大的问题,提出一种基于深度学习的全自动指针式仪表读数方法。首先,引入YOLOv7网络提取表盘区域;其次,采用文中提出的VCA-UNet(VGG16Net,improved skip connections and ASPP based U-Net)网络用于分割刻度线和指针;最后,引入PP-OCRv3网络自动获取仪表量程,并利用角度法确定仪表示数。实验结果表明:VCA-UNet网络的MIoU和MPA值较U-Net网络分别提升18.48%和9.36%,且普遍高于其他经典分割网络,仪表读数的平均相对误差为0.614%,且泛化实验的读数绝对误差相对较小,验证了读数方法的准确性和泛化性。展开更多
当前工业指针式仪表读数过程中存在特殊环境下依赖人工和推理精度低等问题,本文提出一种基于改进U2-Net的指针式仪表读数方法。针对目前仪表识别算法推理精度差和模型参数数量过多的问题,将U2-Net编码阶段的RSU4和RSU5的最深层的两个卷...当前工业指针式仪表读数过程中存在特殊环境下依赖人工和推理精度低等问题,本文提出一种基于改进U2-Net的指针式仪表读数方法。针对目前仪表识别算法推理精度差和模型参数数量过多的问题,将U2-Net编码阶段的RSU4和RSU5的最深层的两个卷积更换成深度可分离卷积,并在每个RSU的编码阶段后加入了ECA注意力模块,使模型更好地关注指针和刻度区域,提高指针和刻度的识别精度。本文在收集到的数据集上进行评估,通过对比实验表明,相较于SegNet、Deeplabv3+及U2-Net方法,本文改进的模型查准率达到94.58%,针对两种量程25 MPa和1.6 MPa的压力仪表读数引用误差达到1.012%,具有较好的性能表现。At present, there are some problems in the reading process of industrial pointer instruments, such as relying on manual work and low reasoning accuracy. This paper proposes a reading method for pointer instruments based on improved U2-Net. Aiming at the problems of poor reasoning accuracy and too many model parameters in the current instrument identification algorithm, the deepest two convolutions of RSU4 and RSU5 in the U2-Net coding stage are replaced by deep separable convolutions, and the ECA attention module is added after each RSU coding stage, which made the model pay more attention to the pointer and scale area and improved the recognition accuracy of pointer and scale. In this paper, the collected data sets are evaluated. Compared with SegNet, Deeplabv3+ and U2-Net methods, the accuracy of the improved model in this paper reaches 94.58%, and the reference error of pressure instruments with two measuring ranges of 25 MPa and 1.6 MPa reaches 1.012%, which has good performance.展开更多
文摘针对光照不均匀和水表表盘雾化的指针式水表在读数检测时出现漏检、误检等问题,提出一种基于改进YOLOv5s的指针式水表读数检测方法。首先,采用Mosaic、Mixup等数据增强方法,提高模型的泛化能力;其次,引入加权双向特征金字塔网络(bilateral feature pyramid network, BiFPN)实现更高层次的特征融合使得水表图像的深层特征图和浅层特征图充分融合,提高网络的表达能力;然后,嵌入卷积注意力机制(convolutional block attention module, CBAM),在通道和空间双重维度上强化指针式水表子表盘示数特征;最后将完全交并比损失函数(complete intersection over union loss, CIoU-Loss)替换为SIoU_Loss(scylla intersection over union loss),提升边界框的回归精度。改进算法的mAP@0.5达到97.8%,比YOLOv5s原始网络提升了3.2%。实验结果表明:该算法能有效提高指针式水表的读数检测精度。
文摘针对现有仪表读数方法易受光照不均等因素影响,而导致读数误差大的问题,提出一种基于深度学习的全自动指针式仪表读数方法。首先,引入YOLOv7网络提取表盘区域;其次,采用文中提出的VCA-UNet(VGG16Net,improved skip connections and ASPP based U-Net)网络用于分割刻度线和指针;最后,引入PP-OCRv3网络自动获取仪表量程,并利用角度法确定仪表示数。实验结果表明:VCA-UNet网络的MIoU和MPA值较U-Net网络分别提升18.48%和9.36%,且普遍高于其他经典分割网络,仪表读数的平均相对误差为0.614%,且泛化实验的读数绝对误差相对较小,验证了读数方法的准确性和泛化性。
文摘当前工业指针式仪表读数过程中存在特殊环境下依赖人工和推理精度低等问题,本文提出一种基于改进U2-Net的指针式仪表读数方法。针对目前仪表识别算法推理精度差和模型参数数量过多的问题,将U2-Net编码阶段的RSU4和RSU5的最深层的两个卷积更换成深度可分离卷积,并在每个RSU的编码阶段后加入了ECA注意力模块,使模型更好地关注指针和刻度区域,提高指针和刻度的识别精度。本文在收集到的数据集上进行评估,通过对比实验表明,相较于SegNet、Deeplabv3+及U2-Net方法,本文改进的模型查准率达到94.58%,针对两种量程25 MPa和1.6 MPa的压力仪表读数引用误差达到1.012%,具有较好的性能表现。At present, there are some problems in the reading process of industrial pointer instruments, such as relying on manual work and low reasoning accuracy. This paper proposes a reading method for pointer instruments based on improved U2-Net. Aiming at the problems of poor reasoning accuracy and too many model parameters in the current instrument identification algorithm, the deepest two convolutions of RSU4 and RSU5 in the U2-Net coding stage are replaced by deep separable convolutions, and the ECA attention module is added after each RSU coding stage, which made the model pay more attention to the pointer and scale area and improved the recognition accuracy of pointer and scale. In this paper, the collected data sets are evaluated. Compared with SegNet, Deeplabv3+ and U2-Net methods, the accuracy of the improved model in this paper reaches 94.58%, and the reference error of pressure instruments with two measuring ranges of 25 MPa and 1.6 MPa reaches 1.012%, which has good performance.
