METHOD OF SEARCHING FOR MACHINE-READABLE ZONE OF DOCUMENT ON IMAGE USING ANN, COMPRISING DIRECT AND TRANSPOSED HOUGH TRANSFORMATION Russian patent published in 2025 - IPC G06V10/22 G06V30/418 G06F7/00 

FIELD: data processing.

SUBSTANCE: invention relates to information technology and specifically to searching for a machine-readable zone (MRZ) of a document on an image. Method of searching for a machine-readable zone of a document on an image using an artificial neural network (ANN), containing direct and transposed Hough transformations, consists in the fact that a compact neural network is used to search for the MRZ in a frame to select pixels belonging to the MRZ, wherein in neural network 2 parallel units are used: a convolutional unit and a Hough-unit, wherein the Hough transformation is applied to the input tensor channel by channel, after which next 3 convolutional layers operate with features in the Hough space, wherein the convolutional layer, the output of which goes to the Hough transformation operator, uses the restricted softsign function as the activation function: f(x)=(x/(|x|+1)), in the remaining convolutional layers, the activation function ReLU: f(x)=max(0,x), wherein symmetric indents are used in convolutions, to obtain a search result for the MRZ, a minimum bounding rectangle search algorithm is applied to the remaining ones.

EFFECT: high accuracy of determining both straight and curved MRZs on photographs.

1 cl, 4 dwg