DARF: direction-aware adaptive routing fusion for cross-modal detection

Abstract In complex scenarios such as low illumination and background clutter, visible-light images often suffer from missing target textures and degraded image quality. Although infrared images can complement target contour and edge information, they lack color and surface texture features. Existing visible-infrared cross-modal detection algorithms still exhibit two key shortcomings. First, the shallow features at the front end of the network lack sufficient ability to extract multi-directional and multi-scale details, resulting in weak fundamental feature representations. Second, most cross-modal fusion methods rely on feature concatenation or fixed-weight aggregation, making it difficult to effectively distinguish shared information from modality-specific complementary information and to dynamically adjust the contribution ratio of visible and infrared modalities according to scene conditions, thereby leading to significant inter-modal interference. To address these issues, this paper proposes DARF, a cross-modal detection algorithm based on YOLOv8, following the idea of front-end basic feature enhancement and back-end dual-modal adaptive fusion. Specifically, two core improvements are introduced: first, a Multi-branch Direction-Aware Feature module (MDAF) is embedded at the front end of the backbone network to enhance shallow feature representations and provide stronger support for subsequent fusion; second, a Dual-Modal Adaptive Routing fusion module (DMAR) is deployed between the backbone network and the detection neck to achieve fine-grained cross-modal adaptive feature fusion. Comparative experiments are conducted on the pixel-aligned LLVIP dataset and the KAIST dataset for cross-scene evaluation. The results show that DARF achieves higher detection accuracy than representative existing methods, thereby validating the effectiveness of the proposed modules and the overall framework.