AUTOMATIC VISUAL PERCEPTION AND TRAINING A VISUAL PERCEPTION ALGORITHM

摘要

According to a method for training a visual perception algorithm comprising an artificial neural network (8), a labeled first set (9a) and an unlabeled second set (9b) of training images are provided. A translation algorithm comprising at least one generative adversarial network (11a, 11b, 12a, 12b) is trained to map the training images of the first set (9a) and the second set (9b) to a shared latent space (10) comprising a set of latent codes, each of them being uniquely associated to a respective training image of the first set (9a) and a corresponding training image (15b) of the second set (9b). The trained translation algorithm and the labels of the first set (9a) are used to train the neural network (8) to infer labels (14) for the second set (9b) from the latent space (10).