The black-box nature of deep learning methods makes it difficult to interpret the results of the analysis and understand the motivations of a particular decision made by the algorithm, and how it relates to particular characteristics of the input data (Zhang and Zhu, 2018). These characteristics represent a strong limitation in many decision-critical applications: expert users, that is the software designers and developers, will require explanation tools that allow them to understand the behavior of the implemented methods for different setting conditions; end users will need access to a satisfactory explanation for the process that led to the decision: if these conditions are not met, a sufficient trust on the automatic system will not be easily achieved, hindering its adoption in real case scenarios.
Indeed, in the deep learning area there are several emerging trends supporting explainability of decisions, like visualization methods, model distillation and intrinsic methods (Ras et al., 2022). However, explainable deep learning is still in its early phase and more developments are needed, since Spain phone number list current solutions of explaining deep learning are seldom enough to achieve explanations helpful in practice, still impeding to deploy and liably exploit deep learning-based solutions, for instance in application domains such as autonomous driving or facial recognition. At this aim, it has to be taken account that according to the particular imaging application, a different purpose and type of explanation will be needed, so that it is not obvious what the best type of explanation metric should be adopted for each particular scenario.
Fundamental aspect for the development of next imaging applications is the availability of large and well designed public datasets. In the last years, there has been the deployment of large datasets in the computer vision area; the most adopted one is probably the ImageNet (Deng et al., 2009), containing over 15 millions labeled images divided into over 22,000 classes. Starting from it, a subset defined ImageNet Large Scale Visual Recognition Challenge (ILSVRC) (Russakovsky et al., 2015) has been derived, spanning 1000 object classes and containing 1,281,167 training images, 50,000 validation images and 100,000 test images.