In salient object detection, many current methods apply the same processing to features across different scales and stages, neglecting the variations in these features and their impact on prediction. High-level features offer rich semantic information but lack precise spatial details, while low-level features provide more detail but often include background noise, complicating accurate saliency target localization and affecting overall model performance. This study introduces a new network architecture, SANet, incorporating a dynamic token selector (DTS) to selectively extract key features and reduce the influence of interfering features. To enhance saliency map generation, we propose a target-focused attention mechanism (TFA), combining query-centered sliding window attention with pooling attention to improve the model’s handling of multi-scale features. This design enhances feature expressiveness and the accuracy of salient object detection. Additionally, we introduce a joint supervision strategy to refine features at each level. Experiments on five challenging datasets show SANet’s superiority over current state-of-the-art methods in salient object detection.