Trigonometric feature learning for RGBD and RGBT image salient object detection

RGB-depth (RGBD) and RGB-thermal infrared (RGBT) images are integral to multi-modal salient object detection (SOD). Despite the impressive results of existing RGBD and RGBT SOD networks, two key areas remain in need of improvement. First, current methods rely heavily on feature fusion techniques, leading to overly complex and inflexible network architectures. Second, these methods lack a unified, adaptive strategy for combining features across different modalities or layers. To address these limitations, we propose a novel Trigonometric Feature Learning (TFL) strategy for generalized feature fusion in multi-modal SOD. Drawing inspiration from the trigonometric principles underlying vector operations, where the dot product of two vectors is the product of their magnitudes and the cosine of the angle between them, our TFL strategy maps features into graph space to compute the "cosine"mapping value for feature fusion. This cosine value dynamically adjusts based on feature attributes, enabling adaptive and effective fusion. To validate the TFL strategy, we design two network structures that use the TFL as the sole feature fusion mechanism for multi- modal SOD. Comparative evaluations against state-of-the-art methods demonstrate the strong performance of our networks, highlighting the unified and adaptive capabilities of the TFL strategy. The source code is available at https://github.com/huanglm-me/TFL-Net.git.