Cross-Collaboration Weighted Fusion Network for RGB-T Salient Detection

RGB-T salient object detection (SOD) aims to accurately and effectively detect and segment salient objects by using the complementary information from RGB and thermal modalities. However, existing RGB-T SOD methods often overlook the fact that RGB and thermal images may not consistently aid in precise detection. To address this issue, we propose a novel approach called CWFNet for RGB-T SOD enhancing the model’s adaptiveness in variable illumination scenarios. First of all, we introduce a global illumination learning module (GILM) to evaluate the illumination conditions. The generated illuminance weights not only adjust modality interaction but also serve to weaken the influence of undesirable modalities. Subsequently, in the encoding stage, guided by the illuminance weights, we design an RGB-guided fusion module (RFM) to complete early fusion of cross-modal features. In the decoding stage, we propose a thermal-guided localization supplement module (TLSM) to strengthen the perception of salient object locations. Extensive experiments are conducted on three benchmark RGB-T datasets and compared to state-of-the-art methods. Our experimental results validate the superiority of our model in the RGB-T SOD task.