AN UNSUPERVISED LEARNING METHOD TO DETECT TRANSPARENT, OR HARD TO SEE, ANOMALIES IN IMAGES

A method of training a neural network to detect anomalies in images comprises, in each epoch, deriving a first input to the neural network by synthesizing an anomalous image from a raw image and inputting the synthesized image to a filter module comprising a filter adapted to increase visibility of anomalies, the filter module applying successive filtering steps, comprising filtering with a first, and then a second, set of filter parameters or using a first type, and then a second type, of filter, and deriving a second input to the neural network by feeding the raw image through the filter module, and feeding the output of the filter module to a loss calculator together with an image output by the neural network, the output image representing a reconstructed image including the predicted location of each anomaly in the synthesized image.