Refined Depth-Quality Feature Processing for RGB-D Salient Object Detection
Ramakrishnan Raman, Amit Gantra, K Kirubasankar, Adusupalle Muni Raju · 2025
Salient object detection (SOD) in RGB-D data relies on integrating depth information with RGB features to highlight the most visually prominent objects in a scene. However, the presence of noisy or low-quality depth maps poses significant challenges to achieving accurate detection. This paper introduces a refined depth-quality feature processing framework designed to improve the precision and robustness of RGB-D SOD. The proposed approach consists of three key components: a Depth Quality Assessment Module (DQAM) to analyze and enhance the reliability of depth maps, a Multi- modal Adaptive Fusion Network (MAFN) to seamlessly combine refined depth features with RGB data, and a Saliency Refinement Network (SRN) for iteratively improving saliency predictions. By addressing the depth quality issue directly, our framework ensures the generation of more accurate and meaningful feature representations. The effectiveness of the method is demonstrated through extensive evaluations on widely used RGB-D datasets, including NJU2K, DUT-RGBD, and STEREO. Experimental results show that the proposed framework achieves superior performance in terms of precision, recall, and mean absolute error (MAE) compared to existing state-of-the-art models. The visual and quantitative results confirm the robustness and reliability of the proposed approach, marking a significant advancement in the field of RGBD SOD and paving the way for further research into quality-driven multimodal object detection.