Number recognition algorithm on the image with a low probability of the second kind of error

There are a number of cases in the industry that require
scene number recognition. Often the required condition for the recognition algorithm is a low value of the second kind of error, namely, cases when an invalid number is recognized. An example of such tasks is:

1. Number recognition on discount, bank cards, Figure 1.
2. Car number recognition, Figure 2.

1 –
2 – ,

, , :

• ;
• ( );
• , , ..

(scene number recognition) : 0.03.

, CRNN (Convolutional Reccurent Neural Network)[1].

github.

Python3, PyTorch.

PSPNet[2]. , github PSPNet Pytorch.

CRNN,
medium [3], [4].

CRNN 3.

3 – CRNN

. , : CNN [5], LSTM [6].

:

1. CNN. . , , , , . , . , , 4;
2. LSTM. LSTM (time step). LSTM . LSTM many to many, . , Bidirectional LSTM, ;
3. . . — ;
4. . n Yn: kn = max(Yn). , , . , , : «3200-544». "-" , . , «00» «44», .

4 –
: h, w — ; n — .

, , 5.

5  –  

, : .

.

CRNN , 6.

  6 – . : , , . CRNN 1, CRNN 2 —

, , . - .

: .
, , . , S = (280, 64), S2 = (320, 64).

, . S = (280, 64), 1.

1 – .
: BS — ; AS — ; k, s, p — , , , : max_pooling

. , . PSPNet.

400 , — 100 , , , 5-10 % , , 5.

2 – . inter_bad — , inter_good — ; good_1, good_2 — , ; amount_cards — , percent_good_1, percent_good_2 — , ; percent_good — ; percent_bad —

, , 1, 0.8816, 0.1184. , - .

, 0.0177, 0.863813, 0.0954 0.0230. , .

, —

, ,

:

• . , . , , ;
• . , ;
• . .

, CRNN scene text recognition, .
CRNN, , .

In addition to this approach, I tried to cut off false predictions with a probability less than a certain threshold, however, in this case, the accuracy of the prediction fell to 0.3, which was unacceptable.

List of sources

1. Original CRNN article;
2. Pyramid scene parsing network
3. Build a Handwritten Text Recognition System using TensorFlow;
4. An Intuitive Explanation of Connectionist Temporal Classification;
5. Convolutional neural network in python
6. LSTM - Networks for Long-Term Short-Term Memory