A Guided Tour of Post-hoc XAI Techniques in Image Segmentation
Résumé
Deep learning models have shown tremendous gains in computer vision in the last decade. However, given their highly non-linear nature, they are often seen as black-boxes. This has led to the development of eXplainable Artificial Intelligence (XAI) as a parallel field with the aim of investigating the behavior of deep learning models. Research in XAI, however, has almost exclusively been focused on image classification models. Dense prediction tasks such as image segmentation have received little attention. The last few years have seen a shift in this trend with works focusing on exploring XAI in the context of image segmentation. A fair number of these works have borrowed from XAI techniques proposed in the context of image classification. It is safe to assume that going forward the number of XAI techniques focused on image segmentation are bound to increase. Reviewing the journey of XAI in image segmentation thus far would therefore be ideal, and is the goal of the present work. This review aims at presenting an overview of the XAI techniques proposed in the context of image segmentation. Another goal is to highlight the lack of interest in this field and its potential causes as well as to comment on potentially underexplored avenues. Given the relative nascency of the field, no review papers currently exist, a gap this work aims to fill.
Origine | Fichiers produits par l'(les) auteur(s) |
---|