XAFNet: Bridging Classification Accuracy and Intrinsic Explainability in Monument Recognition Through Adaptive Dual-Branch Fusion
Year: 2026
Type: article
Abstract: Cultural-heritage AI must justify its predictions to be trustworthy, yet post-hoc explanation methods such as Grad-CAM offer no formal faithfulness guarantee. We propose XAFNet, an inherently interpretable dual-branch network for ancient-Egyptian monument recognition that integrates an EfficientNet-B4 local-texture branch and a ConvNeXt-Tiny global-structure branch through a novel Explainability-Aware Fusion Module (EAFM) with adaptive gating ( $\alpha $ , $\beta $ ), built-in spatial attention generation, and an attention-consistency regularization term. XAFNet is trained with a multi-objective loss that jointly optimizes classification accuracy, attention diversity between branches, and gating–attention consistency. On the Egypt Monuments Dataset v1, XAFNet attains 99.61% $\pm ~0.14$ % mean test accuracy across stratified 5-fold cross-validation — statistically indistinguishable from eight baselines spanning CNN, Transformer, and prior-fusion families, with every pairwise comparison yielding Holm-corrected p > 0.05 under both paired t-test and Wilcoxon signed-rank testing — while uniquely providing intrinsic explanations whose faithfulness is measured directly: the local-branch attention attains an insertion-AUC of 0.559 and a deletion-AUC of 0.376, outperforming post-hoc Grad-CAM computed on the same network against 0.518 and 0.382 respectively on both measures, with attention diversity $0.878~\pm ~0.038$ and a fused feature-space Silhouette of $0.351~\pm ~0.031$ that exceeds either branch in isolation. The contribution is intrinsic explainability at classification parity, not classification superiority. Cross-dataset evaluation on the Architectural Heritage Elements (AHE) benchmark confirms architectural generality, with XAFNet outperforming both baselines at 96.95% $\pm ~0.17$ %, with a fold standard deviation three to four times smaller, and a frozen-backbone linear probe transferring at 93.66% $\pm ~0.24$ %.
Source: IEEE Access
Institution: Helwan University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7214786332