A comprehensive survey of recent XAI papers published in top-tier computer vision and AI conferences and journals (2020-2025)
Explainable AI (XAI) is crucial for fostering human trust in deep neural network (DNN) predictions, particularly in tasks like image classification. Multiple surveys exist on XAI methodologies, however, the practical usability and reproducibility of these methods remain largely unexplored.
This paper addresses this gap by conducting a comprehensive survey of recent XAI papers published in top-tier computer vision and AI conferences and journals (supplemented with a systematic Scopus search). We categorize these works along 7 taxonomy dimensions (including a new "Type" dimension), identify prevalent datasets and evaluation metrics, and analyze the associated code repositories. A novelty of this work is its systematic analysis of the code repositories accompanying surveyed papers. The results are noteworthy: ~95% of codebases are research prototypes, and two-thirds show inconsistencies with their publications. This helps explain why benchmarking XAI methods remains so difficult in practice.
Local explanations focus on specific instances, while global explanations describe overall model behavior.
Post-hoc methods explain trained models; self-interpretable models are inherently transparent.
Deliberative methods justify predictions; counterfactual methods explain "why not" scenarios.
Model-agnostic methods work with any model; model-specific methods are tailored to architectures.
Different technical approaches to generate explanations.
Visual explanations show heatmaps/saliency maps; textual explanations provide descriptions.
Feature-based methods highlight pixels; concept-based use high-level attributes; contrastive/counterfactual compare alternatives; sample-based use prototypes.
These 11 publicly available datasets are the most commonly used for evaluating XAI methods in reviewed papers.
Core set of methods frequently employed in state-of-the-art comparative studies.
Quantitative metrics used to evaluate XAI methods in computer vision.
Analysis of available code repositories for XAI methods, including repository statistics and code quality assessment.
Vahidin Hasić and Senka Krivić
Faculty of Electrical Engineering, University of Sarajevo, Bosnia and Herzegovina
Contact: vahidin.hasic@etf.unsa.ba
This interactive review accompanies the published paper:
A Systematic Review of Explainable Convolutional Neural Networks for Image Classification
Vahidin Hasić and Senka Krivić — Expert Systems, Vol. 43, No. 7, e70279, 2026.
@article{hasic2026xai,
author = {Hasić, Vahidin and Krivić, Senka},
title = {A Systematic Review of Explainable Convolutional Neural Networks for Image Classification},
journal = {Expert Systems},
volume = {43},
number = {7},
pages = {e70279},
year = {2026},
doi = {10.1111/exsy.70279},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.70279},
note = {e70279 5145725}
}