ACSE: Advantage complementation and salient feature-enhanced fusion for visible-infrared object detection

Visible-infrared (RGB-IR) image fusion is pivotal for robust object detection in complex environments, yet existing methods predominantly rely on symmetric architectures and global attention, which fail to fully exploit modality complementarity and to effectively enhance salient features in regions where targets and backgrounds share similar characteristics. This paper proposes the Advantage Complementary and Saliency-Enhanced Fusion Network (ACSE-Net), introducing three innovative and synergistic modules to tackle these challenges: (1) the Advantage Complementary Feature Fusion module, which leverages modality-aware asymmetric extraction (sequential large-kernel expansion for IR contours and parallel polynomial-kernel transformation for RGB textures) combined with dual-dimensional adaptive gating to achieve bidirectional residual complementarity; (2) the Salient Enhancement-Global Collaborative Fusion module, integrating position-aware local salient reinforcement with Flash Attention-driven bidirectional cross-modal interaction to simultaneously bridge global contextual gaps and address local discriminative blind spots; (3) the Differentiated-Weighted Attention Head, employing spatially-prior weighted convolution with hierarchical dynamic parameters and task-adaptive modulation to precisely amplify central responses and jointly optimize localization and classification. Experimental results on multiple public benchmarks, including FLIR, M3FD, and DroneVehicle, demonstrate that our method attains state-of-the-art performance in detection accuracy and model efficiency, surpassing symmetric fusion paradigms and providing a generalizable framework for lightweight multimodal perception with significant practical implications.