Mamba-driven sifter for salient object detection

Attention mechanism, as a key technique for highlighting salient features, has been widely utilized in salient object detection task. However, traditional attention mechanisms, like those in Transformers, which compute attention contexts for tokens and channels en masse, risk diluting salient information, especially in the presence of complex and distracting backgrounds. To address this issue, we employ the selective State Space Model (SSM) in Mamba to alleviate saliency dilution due to the fact that SSM computes attention contexts sequentially for each token and accumulates long-range context through hidden states of preceding tokens. This sequential scanning mechanism selectively filters noise and emphasizes local and historical hidden states, enabling more effective suppression of background distractions and amplification of salient regions. To get a further step for the saliency patterns dilution caused by the channels en masse, we design a channel-wise Mamba sifter to salient object detection. To be concrete, a Mamba-driven channel-split sifter and a channel-merging sifter are composed to emphasize salient attributes within channel groups instead of channels en masse, which helps prevent the drowning of critical information. Experiments on public datasets demonstrate the state-of-the-art performance of the proposed model. Code is available on https://github.com/liuyi1989/MSNet.