MambaAlign: Alignment-aware state-space fusion for RGB-X industrial anomaly detection

Abstract Industrial visual inspection increasingly incorporates complementary sensors, including depth, thermal, and surface normals, to capture defects that Red, Green, Blue (RGB) imagery alone cannot reveal. Current fusion approaches face three limitations that hinder reliable, deployable inspection: convolutional neural networks exhibit limited local receptive fields that impede aggregation of long-range and orientation-dependent context necessary for elongated or subtle defect detection; vision transformers deliver global interactions but incur quadratic compute and memory costs that scale poorly with high-resolution multimodal inputs common in industrial settings; and modest sensor misregistration together with modality-specific noise lead to cross-modal contamination and degraded pixel-level localization. To address these gaps, we propose MambaAlign, an alignment-aware state-space fusion framework that refines each modality with Per-Modal Mamba Modules (PMMs) built on state-space models and QuadSnake scanning to capture long-range, orientation-aware context while preserving spatial coherence. MambaAlign enables semantic, content-conditioned cross-modal exchange through a lightweight Cross Mamba Interaction (CMI) applied at deep semantic stages, which provides cross-modal guidance with near-linear complexity and reduced sensitivity to spatial offsets. A top-down alignment-aware fusion (AAF) reconstitutes low-level channels via local fusion and channel reconstruction, tolerating small spatial misalignments and preserving precise localization. Extensive evaluation on multiple multimodal anomaly detection benchmarks demonstrates large, consistent gains in both image-level detection and pixel-level localization; averaged across three data sets, MambaAlign improves I-AUROC (image-level area under the receiver operating characteristic curve) by 4.8%, P-AUROC (pixel-level area under the receiver operating characteristic curve) by 5.0%, and AUPRO (area under the per-region overlap curve) by 6.5% while maintaining a competitive runtime of 30 frames per second (FPS).