IMAGE PROCESSING METHOD WITH TRAINED NEURAL NETWORKS Russian patent published in 2022 - IPC G06V10/82 G06V20/58 G06T1/40 

FIELD: image recognition.

SUBSTANCE: invention relates to the field of image recognition, namely to a technique for detecting and classifying objects in images using 3D modeling and generating adversarial neural networks. The method comprises the steps of obtaining sets of natural objects and a set of background images, as well as three-dimensional models of objects, by rendering a three-dimensional model of each of the objects with a background, initial sets of synthesized images are obtained using an adversarial neural network consisting of a generator neural network and a discriminator neural network, using a generator neural network, each image is iteratively changed from the original sets of images, using a discriminator neural network, at this iteration, the modified image is compared with a set of full-scale images of objects with a background, the differences are evaluated and the estimated differences are transmitted to the generator neural network, which, taking into account these differences, changes the image, and the estimated differences are used in the discriminator neural network to increase its ability to compare images, and the previously described actions are repeated until the discriminator neural network is able to highlight the differences, and when the possibility of distinguishing the differences is exhausted, the sets of modified images as training data samples are used for training of the respective classifying neural networks, the corresponding trained classifying neural network is applied to the examined image to detect and classify the depicted object.

EFFECT: increasing the probability of detecting and classifying objects on images of various types, including images of the visible range, microwave range and infrared range.

7 cl, 9 dwg