Eliteun team developed a series of algorithms for recognizing the readings of analog meters and digits meters. Embedded in Eliteun smart meter reading module is a low power industrial mini camera which takes photos of the dial plate. The photo is preprocessed first, such as resize, rotation, noise reduction, etc. Then Long Short-Term Memory (LSTM) derived from Recurrent Neural Network (RNN) and other computer vision algorithms are applied to extract information. For pointer meters, the angle represented by the pointer is recognized and calculated to produce the reading.
Image preprocessing is completed mainly via OpenCV. It is a computer vision library that can run on different operation systems such as Linux, Windows, Android and Mac OS. OpenCV is lightweight and quite efficient as its optimized C coding helps to accelerate its operation speed. It supports interfaces for Python, Ruby, MATLAB and other languages, also provides various universal algorithms for image processing and computer vision. Its performance exceeds other major vision libraries, especially in image resize, optical flow and neural nets.
LSTM is an improvement based on RNN, which eliminates the classification issues caused by gradient disappearance in RNN. The design and structure of LSTM is fit for processing events with longer intervals and latency in time series, which helps to yield better accuracy and efficiency than RNN.
The excellent performance and efficiency of OpenCV enables smooth preprocess and excellent quality. The advantages of LSTM are fully leveraged in the scenario of meter reading. Thus Eliteun Hybrid Meter Reading System (HMRS) is powerful, fast and flexibly applicable in reading different meters such as pointer meters or digits meters; and in different industries such as water, electricity or gas; and in different scenarios such as metropolitan centralized meters, or distributed meters in remote areas.