METHOD FOR OPTICAL RECOGNITION OF TEXT LINES BASED ON FULL-CONVOLUTIONAL ANN, QUASI-SYMBOLOGY AND DYNAMIC PROGRAMMING Russian patent published in 2025 - IPC G06V30/10 G06V30/40 G06T1/20 

FIELD: information technology.

SUBSTANCE: invention relates to the field of information technology, namely to the field of document recognition. Disclosed method of optical recognition of text lines on a mobile device from document images, based on full-convolutional ANN, quasi-symbology and dynamic programming, consists in the fact that an input image is obtained, then input image of size W × H to image of fixed height W × H_g, a full convolutional neural network is used to obtain a projection of the image on the horizontal axis P(x) of size W₁ × 1 × A, where A is the size of the recognition alphabet, a projection P1(x) of size W × 1 × A to obtain full compliance between columns of input image and projection, using dynamic programming based on P1(x), constructing a line segmentation path taking into account predetermined restrictions on the width of characters w ∈ [w_min, w_max].

EFFECT: enabling recognition of an image of a text string, without additional data transmission to a server, without restriction of generality on the writing or language that the string contains.

1 cl, 2 dwg