Готово!
Ваша форма отправлена.

Recognition of numbers on the bottle caps on the conveyor

Task:

Build a number recognition system on the caps, at a speed of at least 25 caps per second using an office computer. There are photos of 4 sets of caps of different colors and information that the types of caps, colors and fonts are numerous.

Solution:

After extensive research, we developed a solution based on a cascade of neural networks, convolution and perceptron neural network

The first network locates the digits. the second network specifies the location in the area found from the first network. the third determines the direction of the text. The fourth recognizes the numbers. The first network to spend the most time-consuming work, it takes 10 milliseconds per image (or 100 images per second).

Principal architecture:

Тут будет картинка

Technical specifications of our prototype:

Personal computer:
CPU: intel core i5 (9400), 6 cores, but we use 2 core
GPU: nvidia 1080ti
RAM: 16Gb, but we use 4Gb RAM
SSD disk 1Tb, but for research we use less 100Gb

Soft:
Ubuntu 18.04
Python 3.6.8
Keras 2.2.4 and Tensorflow 1.4
Our reseach code of neural network
Our prepared dataset

Main features our arhitecture:
Our neural network trains the others in a cascade (Neural network teaches neural network).
For new caps, colors, a person without programming can train our neural networks. (Person needs to prepare for the first network up to 10 images and 2 masks for them.)
Person can see the results of work and training network at each level of the cascade. Quick learner.
Our software may be ported to c++ or another GPU arhitecture, but it is not necessary.

DEMO prototype:

We have datasets that weren't trained by the network. You can see the mechanism to work on them. Or you can upload your own cap image.

OR
load you image