Key point localization and recurrent neural network based water meter reading recognition.

Due to the complicated arrangement of the pipes in the narrow space leads to random orientation of the mechanical water meter dial meanwhile its digit wheels are accompanied by arbitrary angle rotation, which makes the detection and recognition of meter reading more difficult. Even the latest visual network technology cannot deal with the challenges. In this paper, two special visual task networks are being closely cooperated to solve above issues. First, a professional water meter detection method is proposed by redesigning and retraining a human joints detection network to accurately locate four key points of reading region. Based on key points the distorted reading region will be geometric corrected by using homography relation to reduce the interference from shooting angle and improve accuracy of subsequent digit recognition. Then, a water meter reading recognition method is proposed by modifying a recurrent block convolutional network. The robustness of digit recognition is improved by block recognition and transcription of reading region features. During transcription stage, we add new recognition markers and probability vectors between each digit in dictionary to solve the issue of digit wheels rotations. Finally, our method achieves more robust water meter detection in harsh environment and higher recognition accuracy. Experimental results showed that our method can get better performance in detection efficiency (6.15 fps) and accuracy (95.30%) compared with recent related works and closer to the level of practical application. • It focuses on the practical living application of water meter reading recognition in complex real scenes. • We redesign and train a human joints detection network to locate the key points of water meter for robust reading region detection. • A recurrent neural network based digital character sequence recognition method is designed to solve robust water meter reading recognition in lossy scene. • Our network can get better performance in detection efficiency (6.15 fps) and accuracy (95.30%) of water meter reading. [ABSTRACT FROM AUTHOR]