Automatic Reading Method for Pointer Meters Based on Improved Deeplabv3+

In response to the problems of imprecise and poor segmentation of pointer meters by the original Deeplabv3+ algorithm in complex environments, this paper proposes an automatic meters reading method based on the improved Deeplabv3+. Firstly, the semantic segmentation algorithm Deeplabv3+ is structurally improved, including replacing the backbone network with Res-Net50, using DenseASPP instead of ASPP and adding a new coordinate feature refinement module to improve the feature extraction and the perception of edge information, so as to extract the dials and pointers of meters more accurately. The experimental results show that the improved Deeplabv3+ prompts MPA by 2.46% and MIOU by 3.59% compared with the original model. These results are better than other segmentation algorithms. Subsequently, the post-segmented image is expanded into a rectangular map by polar coordinate transformation, and the readings are completed using the distance method. It is verified that the average relative error between the adopted automatic reading algorithm and manual reading is 0.25%, meeting the accuracy requirement. The proposed automatic reading method proves its strongly potential in intelligent meter identification.