Development of an automatic reading recognition system for liquid-in-glass thermometers based on deep learning

The reading recognition system based on computer vision technology can improve the reading efficiency and accuracy of liquid-in-glass thermometers (LiGTs), and avoid the error caused by the human-eye reading method. Due to the requirement of a large field of view in the reading recognition process, the targets of LiGTs in the image are small and the accuracy of traditional computer vision (TCV) methods is undesirable. In addition, the potential shadow lines and the meniscus bottom on the image of LiGTs can also affect the robustness of the accuracy of the TCV methods. Around these issues, this paper develops an automatic reading recognition system for LiGTs based on deep learning, which aims to improve the accuracy of automatic reading recognition. Specifically, the cameras in our system are assisted in acquiring ideal images of LiGTs through the back-light illumination and lifting mechanism. Then a multi-task attention network and a general alignment processing module are designed for reading recognition on images of LiGTs. Experimental results on two self-built thermometer datasets prove that the system designed in this study can accurately recognize the temperature reading of LiGTs, and the performance of the proposed method outperforms other reading recognition methods.