关键词: 伪装目标检测, 深度感知, RGB-D融合, 交叉细化, 跨模态学习

Abstract: Camouflaged objects, due to their high visual similarity to surrounding backgrounds in terms of texture, color, and structural patterns, present a formidable challenge in accurate localization and segmentation. RGB-based representations are particularly vulnerable to such interference, often resulting in incomplete structures or even the omission of targets, thereby degrading detection reliability. To address these limitations, an RGB-D Dual-stream Mirror Network (RDMNet) was proposed. A hybrid backbone, composed of TransNeXt and Vision Mamba, was adopted to extract informative features with reduced parameter overhead. To enrich feature representations, Complementary RGB and depth signals were leveraged by a Multi-modal Feature Fusion (MFF) module through residual integration, enhancing depth cues for improved boundary inference. Further refinement was achieved via a Depth Positioning Module (DPM) and a Positioning-Guided integrity Aggregation (PGA) module. Precise contour-aware localization was provided by the former, while object identification was accelerated and segmentation completeness was reinforced by the latter through iterative cross-modal enhancement. Both global structural coherence and fine-grained boundary precision were jointly emphasized by these components. To capture subtle textural variations, particularly within shallow feature layers, a Convolutional gated Channel Attention (CCA) module was introduced to selectively emphasize structurally relevant information. Empirical evaluations on COD and RGB-D SOD benchmarks confirm the superiority of the proposed framework over 15 state-of-the-art methods. On CAMO, COD10K, and NC4K datasets, RDMNet achieved consistent gains over Multi-View Guided Network (MVGNet), with average improvements of 1.9% in structural similarity index measure (s-measure), 4.7% in mean enhanced alignment measure (mean e-measure), 3.1% in weighted f-measure, and a significant 17.5% reduction in mean absolute error. These findings highlight the model’s effectiveness in enhancing both segmentation completeness and precision in camouflaged object detection.

Key words: Camouflaged Object Detection (COD), depth awareness, RGB-D fusion, cross-refinement, cross modal learning