A coarse-fine reading recognition method for pointer meters based on CNN and computer vision

To enhance the robustness and remove the accumulative error of existing methods, this paper proposes a novel coarse-fine pointer meter reading recognition approach using CNN in the whole recognition procedure. Firstly, the Mask R-CNN is employed to localize the dial position of a meter. Secondly, the dial center is determined by using all the digital scale regions recognized by the R-CNN, while the pointer is extracted by using the regional growth method. The meter's rough reading is then accomplished according to the position of the pointer and its two closest scale marks found by circular scale searching. Finally, the meter's exact reading value is recognized by using the proposed CNN model. A set of reading recognition experiments on various meters, meters with disturbances, and on-site meters have been conducted to verify the proposed approach. The experimental results show that the proposed method is robust under various environments and its maximum fiducial error in all the experiments is 0.63%, which is less than the error of the existing methods.