Robust water meter reading recognition method for complex scenes

Abstract

Abstract We propose a robust water meter reading recognition method for complex scenarios. In our method, the traditional water meter recognition method has been improved so that it can handle more variable real-world scenarios. A WDPDet network is applied in the digit and pointer area detection. The geometric methods is used for rotation correction for subsequent processing. In particular, a data augmentation method is proposed for half-word character recognition that is difficult to handle in the past, which greatly improve network performance. A coarse-grained projection method for pointer recognition is proposed to deal with more complex interference. We have performed a lot of experiments on different devices and scenarios, and have achieved good accuracy.