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.
Automatic Pointer Instrument Reading Method Based on VCA-UNet
Aiming at the problem that the existing instrument reading methods are easily affected by unequal illumination fac-tors,which leads to large reading errors,a fully automatic pointer instrument reading method based on deep learning was pro-posed.First,the YOLOv7 network was introduced to extract the dial area.Secondly,the VCA-UNet(VGG16Net,Improved Skip Connections and ASPP based U-Net)network proposed in this paper was used to separate the scale and pointer.Finally,the PP-OCRv3 network was introduced to obtain the instrument range automatically,and the instrument representation number was de-termined by angle method.The experimental results show that the MIoU and MPA values of VCA-UNet network are 18.48％and 9.36％higher than those of U-Net network respectively,and are generally higher than other comparison networks.The average relative error of meter reading is 0.614％,and the absolute error in the generalization experiment is relatively small,which verifies the accuracy and generalization of the reading method.
NETL
NSTL
万方数据
