Computer Vision News - August 2021

Steffen Czolbe 21 function s, which compare images during training. More recently, these have been superseded by networks trained on image classification – using VGG, for example, to extract features for comparison . This idea of using a separate network to extract features and then judge the distance of images based on that was what inspired this work. However, an image classification network trained on ImageNet may work well when you generate faces because the datasets are somewhat similar, but medical images are radically different, with a range of different modalities . To overcome this, Steffen trained specific networks to make sense of these images. He explored using both an autoencoder and a segmentation mode l – the autoencoder has the benefit of being unsupervised, while the segmentation network needs segmentation masks to train. Looking to the future, Steffen tells us he is currently working on a journal version of this work which has already moved it on further. “There have been some developments around other baselines I want to compare against, ” he reveals. “For example, mutual information is a metric that traditionally people use on image registration, but when I first set out, it hadn’t really been adapted to deep learning. Now, I’m looking forward to comparing against that. Overall, I want to discuss the different similarity measures in more detail .” Best of M I D L