HCMFM: hierarchical cross-modality fusion module for object detection

Detecting targets under adverse visual conditions—including darkness, haze, and heavy occlusion—remains unreliable when the system depends on a single sensing source. Visible-spectrum imagery offers abundant color and texture cues, yet its reliability rapidly deteriorates under poor lighting. In contrast, thermal imagery remains stable across illumination changes but sacrifices detailed appearance information. To leverage the heterogeneous strengths of both sensing streams, we introduce a structured hierarchical fusion architecture termed the Hierarchical Cross-Modality Fusion Module (HCMFM), for robust RGB–infrared object detection. The proposed module operates at the representation level and can be inserted between backbone and head components without altering the original detection pipeline and enables multi-scale cross-modality interaction in a unified manner. Specifically, it consists of three cooperative components: a Cross-Modality Fusion Block (CMFB) that captures global cross-modality dependencies through transformer-based token fusion, a Cross Attention Interactive Block (CAIB) that explicitly mines complementary information between modalities via cross-attention mechanisms, and an Adaptive Gate Fusion (AGF) module that dynamically balances different fusion pathways according to scene characteristics. This hierarchical design effectively enhances discriminative representations while suppressing cross-modality interference under challenging conditions. Results obtained from standard benchmark datasets show consistent performance gains of the proposed framework relative to existing RGB–infrared detection approaches.