Salient Object Detection (SOD) is a difficult task focused on accurately identifying and segmenting prominent objects within an image. Traditional methods often fail to effectively balance global and local contexts, capturing either broad structures or fine details. This paper introduces RMFNet, a Recursive Interactive Multiscale Feature Fusion Network that unifies global semantic understanding with local detail refinement. The proposed method adopts an encoder-decoder architecture, where the encoder combines Transformer and CNN to better extract global semantic features and local detail features. Considering that the features extracted by Transformer and CNN are different, the fusion of features from two branches may introduce noise, so it is necessary to handle them differently during the fusion process. To address the problem, we propose a Multiscale Fusion Module (MFM) that progressively integrates features from both branches in a step-by-step manner. The decoder comprises three hierarchical sub-decoders, each capable of normalizing multi-level feature maps to a unified spatial resolution via recursive pooling and upsampling operations, thereby facilitating holistic full-scale feature integration. The SOD experimental results on five public datasets show that our approach has better performance under different complex scenes compared with some state-of-the-art methods.