DFSMamba: Detail-Guided and Direction-Adaptive State-Space Modeling for Structure-Preserving Visible-to-Infrared Image Translation

Infrared imagery supports low-light target perception, but acquiring, registering, and annotating paired visible–infrared imagery is costly. Visible-to-infrared image translation can alleviate this demand, yet cross-modal differences often degrade object boundaries and spatial layouts, especially for small targets. We propose DFSMamba to preserve geometric structure throughout visible-to-infrared translation by connecting frequency-guided feature formation, input-dependent directional aggregation, and infrared-domain reconstruction supervision. A discrete wavelet transform (DWT)-based detail-guided module supplements bottleneck features with learned structural cues, direction-adaptive 2D selective scan adjusts the fusion of aligned directional responses, and state-space similarity matching constrains the output in a fixed multiscale feature space learned through real-infrared reconstruction. Experiments on FMB, VEDAI, and AVIID-1 include image-quality comparisons, target-level evaluation, ablation studies, efficiency analysis, and downstream vehicle detection. Among evaluated methods, DFSMamba achieves the best SSIM, MS-SSIM, PSNR, and L1 values on all three datasets. On AVIID-1, it obtains an SSIM of 0.985 and a PSNR of 41.799 dB. A YOLOv9-C detector trained on generated images reaches an AP50:95 of 0.636 on real infrared test images, compared with 0.640 when trained on real infrared images. These results show that DFSMamba preserves detection-relevant structure and supports the use of generated images as synthetic training data in the evaluated aerial vehicle-detection setting.