RP-UpsNet: Residual Projection Fusion With Guided Upsampling for Salient Object Detection

Abstract

Salient Object Detection (SOD) aims to identify visually salient regions by mimicking human visual attention mechanisms. Despite significant advances in deep encoder-decoder architectures, preserving fine spatial details and accurately detecting small or low-contrast objects remains challenging, particularly due to gradual spatial downsampling. This issue leads to degraded edge quality and structural inconsistencies, particularly in complex scenes and non-central regions. To address these limitations, we propose RP-UpsNet, a new SOD framework that integrates guided upsampling and multi-scale feature fusion to reduce the loss of spatial detail. The framework uses the Compact Multi-Scale CNN (CMS-CNN) backbone and distributes high-resolution, content-sensitive features across multiple network stages. In addition, an additive gate-based fusion mechanism effectively combines semantic and spatial information, while the residual projection module enhances cross-scale feature interactions. Extensive experiments conducted on eight different benchmark datasets demonstrate that the proposed method achieves competitive and consistent improvements in $F_{\beta } ^{w}$ , $S_{\alpha }$ , and MAE. The proposed approach demonstrates significant improvements in boundary preservation and structural consistency, particularly in challenging scenarios involving small objects, low-contrast regions, and complex backgrounds. The code will be released at https://github.com/Sbl-Kcd/RP-UpsNet