METHOD FOR NEURAL NETWORK RECOGNITION OF HANDWRITTEN TEXT DATA ON IMAGES Russian patent published in 2025 - IPC G06V30/10 G06V10/82 G06V10/75 G06V30/244 G06F17/00 

FIELD: computer engineering.

SUBSTANCE: invention relates to computer engineering and specifically to methods of recognizing handwritten text using neural networks. Method for neural network recognition of handwritten text data on images is that an image f of a text field of a document is input, the dimensions of which are known, wherein the image f in the RGB colour space containing the text field of the document is further considered, wherein the input image f is processed by the neural network method according to the following algorithm: the image f is supplied to the image preparation unit, where it is converted into a single-channel, and also reduced to height H_f; at the second stage, the pre-trained neural network extracts visual features from the prepared image, forming a feature map M_f, prediction unit with an attention mechanism iteratively decodes a feature map M_f, predicting symbols one by one based on previously predicted ones, wherein if the stop criterion is achieved or the number of iterations exceeds N, then the recognition ends, wherein the extraction of visual features and the prediction unit with an attention mechanism are components of the same neural network model.

EFFECT: high speed of processing, as well as the ability to process input images of arbitrary width.

4 cl, 5 dwg, 1 tbl