BIDP: Brain-Inspired Dual-Process CNN-Transformer for Salient Object Detection

While most existing salient object detection (SOD) methods achieve accurate predictions of salient regions, their performance in precisely delineating object boundaries remains suboptimal. Inspired by the brain’s dual-process mechanism, we propose a novel dual-process framework: (1) a Pyramid Efficient Channel Attention U-Net (PECA-U-Net) that enhances local detail representation during initial salient feature extraction, and (2) a Residual Efficient Channel Attention Transformer (RECA-Transformer) that refines edge details through global context modeling. These modules are sequentially integrated within a feature extraction and refinement backbone architecture, enabling comprehensive boundary-aware saliency prediction. Furthermore, we design a Boundary-Aware Attention Loss (BAAL) that adaptively balances global saliency and boundary precision through a weighted combination of base and boundary losses. Extensive experiments conducted on five benchmark datasets demonstrate the effectiveness and competitiveness of our proposed model in SOD tasks. The source code is publicly available at https://github.com/1245179435/PECA-REAT.