RGB-depth salient object detection technology based on multi-level features and adversarial attacks

Abstract

Salient object detection in complex scenes suffers from insufficient cross-level feature interaction and vulnerability to adversarial perturbations. Existing RGB-D methods mainly focus on feature fusion but lack systematic adversarial defense. To address these issues, a robust RGB-D salient object detection framework integrating multi-level features and joint adversarial defense is proposed, with two main contributions: an Enhanced Detail Integration Module and an Enhanced Semantic Integration Module are constructed with channel-spatial attention to achieve hierarchical fusion of detail and semantic features, and a Contextual Correlations Aggregation Denoising Active Defense Module is embedded together with WebP compression preprocessing to form an active-passive joint defense against typical adversarial attacks. The method was evaluated on the NJU2K, NLPR, and RGBD135 datasets. On clean images, the model achieved an MAE of 0.014 and an S m of 0.971 on NJU2K, outperforming DMRA (0.017, 0.963), an S m of 0.924 on NLPR exceeding DPANet (0.915), and an MAE of 0.030 with an F β of 0.918 on NJU2K, surpassing D3Net, VST, SMTNet, and SPNet. Under multi-type adversarial attacks, the model maintained the best precision of 0.875 under PGD, an adversarial F β of 0.592 ± 0.032 on RGBD135, and an adversarial MAE of 0.098 ± 0.006 on NLPR, with a detection efficiency of 62.5 frames per second. The results verify the superiority of the proposed method in detection accuracy and adversarial robustness, providing technical support for autonomous driving, video surveillance, and medical target diagnosis.