Salient object detection is one of the hot research topics in the field of computer vision. However, when handling images with complex backgrounds, salient object detection often faces the issue of blurred object boundaries. A multi-scale feature fusion network structure, named U²Net_MFF, was proposed. The network introduced skip connections at each decoder layer, allowing each decoder to not only obtain the features from the previous decoder but also acquire feature information from all shallower layers above. This approach enabled the network to capture more detailed features of the target region, thereby improving detection accuracy. Additionally, dynamic upsampling (DySample) was incorporated into the residual U-shaped blocks to enhance the quality of the upsampled feature maps. A multi-dimensional collaborative attention mechanism was also employed to suppress interference from complex backgrounds, while increasing sensitivity to the salient object regions. Experimental results show that the improved algorithm achieves increases of 1%, 0.3%, 0.4%, 0.7%, and 0.8% in accuracy, precision, recall, mean intersection-over-union (MIoU), and F₁ score, respectively. These results demonstrate that the algorithm has stronger anti-interference ability and produces clearer object boundary detection.