A Study of Salient Object Detection Based on Multilevel Interaction

Salient Object Detection (SOD) is designed to identify and segment the most significant objects or regions in an image. Most existing SOD methods typically use U-Net or feature pyramid structures to simply aggregate features at different scales, which often leads to a mixture of low-level noise and a lack of high-level details, ultimately resulting in degraded model performance. In response to this issue, this paper proposes the Saliency Multi-Dimension Attention Network (SMDA-Net) for significant object detection. Firstly, the Multilevel Interaction Module (MIM) is proposed, which introduces a cross-attention mechanism to facilitate the interaction between features at different levels, enabling high-level features to guide the learning of low-level features and enhance salient regions. Secondly, to address the problem that Transformer-based SOD methods tend to ignore the details of complex objects, the Hybrid Attention Module (HAM) is designed, which combines global self-attention and windowed self-attention to model the context at both global and local levels, thereby improving the accuracy and detail retention of the prediction map. Finally, a multilevel supervision strategy is adopted to optimize the decoding process in stages. Experimental results show that SMDA-Net outperforms current state-of-the-art models on challenging datasets, providing a more effective method for significant object detection and being of great significance to the development of this field.