Hello, Habr! Today we continue our classic series of articles on how to use something practical and practical using image processing and pattern recognition methods. Today we will focus on the task of recognizing price tags. The usual price tags for products that anyone can find in any store. Only to ensure proper practicality, we will add an important requirement to the statement of the problem: recognizable images are obtained using a small format digital camera, and the computing device has significant resource limitations. In other words, we’ll tell you how to recognize price tags on a computationally weak mobile device (by the way, this is not so much about a cheap Chinese smartphone, but about special industrial data collection terminals,which, for a number of reasons, also have rather weak “brains”). So, if you are interested in the topic of retail automation, welcome to cat!
, , , ().
, - , . , . , . .
, . , - ( ), .
, . , : Odroid XU4 ( , Honeywell) . retail-, .
:
- .
- () .
- .
- .
- , . : , , .
4 .
:
- : - .
, , .
, , . ,
(, ) (, , -, , , ..), .
- . , . : , , , - .
, . - .
, . , .. .
. .
, . , .
. , . , , .
, . , , , .
. , , . , . , , — -- .
, , (. ) — .
, : .
8- . .
, . , . , , .
, , , .
, . , . , , , , , . , (- ), , , , .
. , . () .
, . , , . , , . , . . .
, .
, .
, .
708 , 29 ( ). 679: ~80 , ~90 , ~150 , ~ 40 50 . 1350x700 800x400.
, .
:
- True Positive (TP) — , ,
- True Negative (TN) — , ,
- False Positive (FP) — , , , - () ,
- False Negative (FN) — , .
:
:
:
, . FP , , TN , FP.
: — 162 Odroid-XU4 .
, . , . — , ( ). , , - // ( ), “ ”. “” ( ) . , , , “ ” , . .