INTERACTIVE IMAGE SEGMENTATION Russian patent published in 2025 - IPC G06T7/12 

FIELD: data processing.

SUBSTANCE: invention relates to a method for interactive image segmentation based on a sequence of user clicks performed by a trained neural network for interactive image segmentation. Method comprises steps of: feeding to the input of a neural network an image, coordinates of a user click; predicted segmentation mask is obtained at the output of the trained neural network, the predicted segmentation mask is displayed to the user. At the training stage, said neural network is trained by means of training coordinates of a click on the training image until the error between the reference segmentation mask and the predicted segmentation mask on the hold-out sampling decreases, wherein the training coordinates of each click are optimized based on the back propagation of the error between the reference segmentation mask and the predicted segmentation mask at fixed weights of the neural network, obtaining the training adversative coordinates of the click; based on the obtained training adversative coordinates of the click, the neural network is trained with cancelled fixation of its weights by backward propagation of the error through the neural network to update its weights.

EFFECT: improved quality and high accuracy of reproducing a segmentation mask predicted by a neural network.

7 cl, 7 dwg, 2 tbl