DEVICE AND METHOD FOR TRAINING A NEURAL NETWORK

Abstract

A device and a method for training a neural network are disclosed, wherein the method of training a neural network includes: a first image processing method generating a plurality of first segmentation images for a plurality of digital images (104), wherein each digital image of the plurality of digital images (104) includes at least one object (202) and wherein each first segmentation image of the plurality of first segmentation images includes a segmentation of the at least one object (202) for the associated digital image; training a first neural network portion (208) using the plurality of digital images (104) and the associated first segmentation images; the trained first neural network portion (208) generating a plurality of second segmentation images (1002) for the plurality of digital images (104), wherein each second segmentation image of the plurality of second segmentation images (1002) includes a segmentation of the at least one object (202) for the associated digital image; a second image processing method generating a plurality of third segmentation images for the plurality of digital images (104), wherein each third segmentation image of the plurality of third segmentation images includes a segmentation of the at least one object (202) for the associated digital image; training a second neural network portion (608) using the plurality of digital images (104) and the associated third segmentation images; the trained second neural network portion (608) generating a plurality of fourth segmentation images (1004) for the plurality of digital images (104), wherein each fourth segmentation image of the plurality of fourth segmentation images (1004) includes a segmentation of the at least one object (202) for the associated digital image; and training a third neural network portion (1008) using the plurality of digital images (104) and the associated second segmentation images and fourth segmentation images.

IPC Classifications