Показания счетчиков
Счетчик газа yesco
Фрисби logo

Neural network for recognition of meter readings

01/

Main information

Let's remember the classic methods of taking readings.
Every Russian once a month pays for the readings of housing and communal services from his analog counter.
FRISBI decided to facilitate this issue and considered several solutions.
Decision
Re-equip apartments with automatic meters
»
«
Expensive
(1,000,000 ₽/house)
Difficult to maintain
Not scalable
Decision
To teach a mobile application to read through the camera of housing and communal services counters
»
«
Мобильное приложение считывает данные ЖКХ счетчиков

Interesting fact!

01
02
This is the first project of its kind in Russia
The neural network recognizes numbers in complex cases better than people
02/

What difficulties did you face?

Semi-visible digits on analog counters
Problem
There was one more subtlety that we managed to cope with. The fact is that the counters are scrolled, and two numbers are visible in the window at the same time
With the help of labeled data, we trained the model for such cases - now it selects a number that is larger in value
Decision

Neural network speed on mobile phones

On mobile devices, the neural network requires simpler solutions due to the insufficient capacities of different devices.
Problem:
Our developers have revised the code for a mobile device without losing the efficiency of the neural network.
Decision:
03/

System operation diagram

01

Loading an image

Аналоговый счетчик
02

Counter Recognition

Распознавание аналогового счетчика
03
Recognition of numbers on it
Распознавание цифр аналогового счетчика
04
Analysis
Анализ аналогового счетчика
04/

Real examples

Counter localization

Счетчик воды
Нейросеть распознаёт счетчик воды
Нейросеть анализирует счетчик воды

Dial localization

Dial alignment

Циферблат счетчика воды
Нейросеть локализирует циферблат счетчика воды
Циферблат счетчика воды
Нейросесть выравнивает циферблат счетчика воды
Нейросесть выравнивает циферблат счетчика воды
05/

Project team

Фрисби logo
Techlead
Konstantin Kubrak
Developer
Viacheslav Schultz
06/

Emotions from the project

Konstantin Kubrak
Manager, came to the project at the "Black Pearl"
Viacheslav Schultz
Developer, thought most of all
It's great that our system has learned to recognize the dials of counters of any type.
The whole team is delighted, because we are working with a completely new product and with a complex modern project, which has no analogues in the CIS
07/

Our plans

Idea
Optimize the neural network for mobile operating systems:
Android
iOS
iOS logo
Android logo
For what
So that the model can work faster on older phones and weaker phones
Contact us