Microsoft Word Avtoreferat (Muhiddinov M. N. Tatu)

«Development of text extraction and 
recognition method and algorithm based on neural network»
, text detection 
and recognition methods based on fully convolutional neural network (FCN) and 
Tesseract OCR model were developed. The proposed text detection and 
recognition methods divided into four main stages as shown in Figure 2: 1) pre-
processing stage global contrast enhancement using histogram equalization method 
2) text detection using FCN, 3) text extraction and recognition, and 4) Uzbek 
language TTS synthesizer. The end-to-end text recognition approaches consist of 
two parts: text detection and text recognition. It should be noted that, existing 
approaches, both conventional and deep neural network-based, principally consist 
of many steps and parts, which are reasonably sub-optimal and time-consuming. 

---

29 
Accordingly, the efficiency and accuracy of such approaches are still far from 
sufficient.
To overcome these drawbacks, a quick and reliable scene text detection and 
localization method that has only two steps proposed. The proposed method uses 
FCN model that instantly generates word or text-line level prophecies, apart from 
unnecessary and heavy intermediate stages. Using convolutional neural network 
(CNN) is not an effective solution to determine the precise location of texts in the 
image. Therefore, the use of the FCN model will yield higher accuracy results.
Figure 2. The proposed text detection and recognition methods and Uzbek 
language TTS synthesizer 
Furthermore, different circumstances must be considered when creating 
neural networks for text detection. Because the areas of text regions differ 
remarkably, discovering the presence of long sentences would need features from 
late-stage of a neural network, while predicting correct geometry surrounding a 
short text regions demand low level knowledge in early stages. For that reason the 
network must utilize features from various levels to satisfy these demands. The 
proposed method slowly unites feature maps while preserving the up sampling 
features merging small. Simultaneously the method concludes with a network that 
can both use various levels of features and retain a small calculation cost. The 
model can be decayed into three parts: feature extractor, feature-merging and 
output layer. The feature extractor might be a convolutional network pre-trained on 
ImageNet dataset, along with interleaving convolution and pooling layers. Four 
levels of feature maps, represented as , are obtained from the feature extractor, 
whose sizes are 
1/32
, 
1/16
, 
1/8
and 
1/4
of the input image, respectively. 
Mathematically, feature-merging formulation expressed as: 
=
(ℎ )
3
×
(ℎ ) = 4
(5) 
ℎ =
= 1
×
(
×
(
;
)) ℎ
(6) 

---

30 
where is the merge base, and 
ℎ
is the merged feature map, and the operator [;] 
denotes concatenation with the channel axis. 
In the next stages, once the text region is detected, the region can be cropped 
and processed further to recognize the text. To do this, trained Tesseract OCR 
model with Uzbek Latin and Cyrillic alphabet characters can be used. The 
proposed method also includes recognized texts send to TTS synthesizer for Uzbek 
language.
In this chapter of the dissertation, the result of studying and analyzing words 
in the Uzbek dictionary, an electronic database of 31,5 thousand words was formed 
and arranged in alphabetical order. The Uzbek language speech synthesizer is 
based on the concatenation method and contains pronunciation of the words. 
Therefore, the Uzbek vocabulary with 31,5 thousand words were studied and all 
words were broken down into 2,5 thousand sections, i.e. syllables. For correct 
pronouncing of recognized texts and update Uzbek language database, recognized 
texts are compared with database, if recognized text is exist in Uzbek language 
database system send it to Uzbek language TTS Synthesizer, else the word send to 
language specialist to confirm new word. 
In the fourth chapter of the dissertation,