Robust salient object detection based on triple attention-guided multi-resolution fusion and feature refinement.

Abstract

Salient object detection (SOD) is dedicated to highlighting the critical image elements in complex visual scenes. However, this task faces two serious challenges: first, salient objects are often submerged in cluttered backgrounds and are susceptible to disturbance from background noise; second, the substantial scale variation among these objects presents considerable challenges for accurate detection. In order to address these challenges, an attention-based method for salient object detection is proposed. Firstly, we innovatively design a Triple Attention-guided Multi-resolution Fusion (TAMF) module, which integrates a spatial, channel, and global attention mechanism to dynamically adjust feature weights to suppress background noise. At the same time, it introduces a multi-resolution feature fusion framework to enhance cross-scale interactions. Secondly, we propose the Feature Refinement (FR) module, which utilizes four parallel convolutional branches and different-scale dilated convolutions in conjunction with the triple attention mechanism to precisely detect and enhance the salient object features, as well as effectively address the challenges of scale changes. Evaluations across five challenging benchmark datasets demonstrate notable improvements over advanced methods, highlighting our model's effectiveness and competitive advantage. Code is available at: https://github.com/zbbany/ATMF_FRNet.git.