<p>This paper puts forward a novel multi-scale Salient Object Detection (SOD) model (SFENet) based on the specific features of enhanced RGB images and depth images. The SFENet model mainly contains two parallel single-modal specific feature learning networks for RGB images and depth images, and a cross-modal shared learning network. For the specific encoders of RGB image and depth image, a granular feature extraction module (KGM) is designed. It uses the K-means++ algorithm to cluster images and divides the images into multiple local regions, achieving the goal of discrete image processing. For the shared encoders, a unified cross-dual attention fusion module (CDM) is proposed. It utilizes the channel attention and coordinate attention to realize cross-modal fusion of specific features of RGB image and depth image. For the shared decoders, a multi-branch feature fusion module (MBF) is presented. It achieves shared feature enhancement by integrating specific features of RGB images and depth images into a shared decoder. Meanwhile, the result generated by the shared decoder is the final result of significance detection. The highest accuracy and average accuracy of SFENet model reaches to 95.17 and 93.22% on 5 RGB-D significance target detection benchmark datasets. Compared with other models, the proposed model has high stability and robustness in complex visual scenes.</p>