Reading Digital Numbers of Water Meter with Deep Learning Based Object Detector

Automatically reading water meter is a classical OCR problem, typical method includes four major components: region of interests (ROIs) detection, skew correction of bounding boxes, single digital character segmentation, and digital classification. Disadvantage of the traditional method is that the pipeline is too complex and coupled to the accuracy of the final recognition result. Deep learning based object detection has achieved promising results on many computer vision tasks. As one of the representatives of the deep learning object detection framework, YOLOv3 perform detection task quickly and accurately. Inspired by this, we formulate the water meter reading problem as a detection problem, which is a true end-to-end solution. In order to attack the half-character problem of water meter, we proposed a heuristic rule to guarantee that there is only one bounding box in the vertical direction within a grid. Experimental results on our own built XMU-W-M dataset showed that the 0-error recognition rate reaches 96.67% and the 1-error recognition rate is up to 99.81%, which outperforms the traditional water meter recognition system in both time and precision. Both the code and dataset are available: https://github.com/sloan96/water-meter-recognition .