Attention-Guided Channel-Spatial Feature Recalibration for Domain-Adversarial Image Classification

Abstract

A Domain-Adversarial Neural Network (DANN) aligns global features but does not control how an added attention block changes the shared representation. We test attention-guided DANN (AG-DANN), which places a Convolutional Block Attention Module (CBAM) before global pooling and scales its residual with a zero-initialized signed gate. With fixed final checkpoints and three seeds, AG-DANN scores 81.68% on Office-31 and 60.87% on Office-Home; DANN scores 81.69% and 60.84%. The complete module does not improve on DANN. In a 72-run Office-31 ablation, gating recovers 4.54 points lost by ungated CBAM, although paired t and Wilcoxon tests give different conclusions after Holm correction. Gate values remain small and change little late in individual runs. Feature and attention diagnostics show no measurable separation or localization gain. Under this protocol, the gate limits damage from direct attention insertion rather than improving DANN itself.