Cognitive Systems and Signal Processing

Fuchun Sun,Huaping Liu,Dewen Hu

DOI: https://doi.org/10.1007/978-981-13-7983-3

2018-01-01

Abstract:Analog pointer instruments are widely used in industrial automation. However, reading their values by human is time and labour consuming. Most current pointer reading algorithms lack universality and can not handle complex and varying application scenarios. Therefore, this paper proposes a general pointer instrument reading scheme based on CNN. Our main contributions are two folds. First, we combine two sub-networks into one backbone, which has shortened the inference time. Second, we propose a light-weighted sampling decoder, which makes our scheme achieve a higher accuracy. Comparing to other algorithms, our scheme can deal with the case in which the instrument is too difficult to read, e.g., there is more than one similar pointer in one panel. Also, our scheme only needs less than 100 ms on a typical laptop for a 512× 512 image. It is fast enough and can satisfy most requirements in industrial automation.