PiSOD: a plug-and-play pseudosupervised framework for semantic-structural alignment in salient object detection

Salient object detection (SOD) simulates the human visual attention mechanism. This task focuses on automatically identifying the most visually salient regions or objects in an image. Although existing methods have achieved notable accuracy, they still rely on costly pixel-level annotations and lack sufficient semantic and structural constraints, particularly the effective integration of pixel-level details with object-level semantic and structural information, which often leads to blurred boundaries, semantic drift, and overfitting under limited samples or noisy labels. To address these issues, we propose a pseudosupervised and integrable SOD framework (PiSOD) that generates fused pseudolabels both semantically consistent and structurally complete, enabling the model to learn high-quality saliency representations under weak supervision. Furthermore, we introduce a Saliency Prior Alignment Module, seamlessly integrated into training, which provides boundary constraints via a structural prior branch, semantic guidance via a semantic prior branch, and a pseudolabel selection mechanism for effective cross-modal alignment, thus enhancing stability and boundary precision. The PiSOD framework supports two modes: fine-tuning existing SOD models to boost accuracy and robustness, and zero-shot inference that directly utilizes fused pseudolabels for efficient saliency prediction. Experimental results demonstrate that, compared with the baseline SOD models, our PiSOD-enhanced approach achieves notable improvements across multiple challenging datasets. Specifically, the integration of PiSOD results in an average mean absolute error (MAE) reduction of about 0.3% and an average improvement of 1.3% in the S metric across ECSSD, HKU-IS, DUT-O, PASCAL-S, and DUTS-TE. Moreover, these improvements are particularly evident in preserving structural integrity and semantic consistency, highlighting the effectiveness of our method in enhancing saliency detection without additional supervised training.