Abstract
Salient object detection (SOD) plays a crucial role in computer vision by mimicking human visual attention to identify and segment the most prominent objects in a scene. The advent of depth sensing technologies has propelled RGB-D SOD to the forefront, leveraging three-dimensional geometric information to enhance perceptual capabilities in complex scenes. This paper introduces a Multi-modal Interaction and Fusion Network (MIFNet), designed to address key challenges in RGB-D SOD, including modal heterogeneity, inadequate multi-scale context, and insufficient deep feature interaction. MIFNet adopts a dual-stream ResNet-34 backbone, integrating a Cross-modal Adaptive Attention Fusion (CAAF) module and an SE-enhanced Multi-modal Feature Interaction Block (MFIB_v2). The CAAF module captures cross-modal contextual information through a hybrid dilated convolution strategy and a dual-channel attention mechanism. The MFIB_v2 module enhances channel interaction and adaptive feature fusion through a three-stream learnable weighting strategy. A systematic four-stage ablation study with 15 comparative experiments was conducted to comprehensively quantify the contribution of each module and their synergy. Evaluations on three public datasets show that MIFNet achieves outstanding performance, with a Mean Absolute Error (MAE) of 0.0623, representing a 2.8% improvement over state-of-the-art methods, alongside significant enhancements in S-measure and E-measure. For a custom industrial weld defect dataset, MIFNet attains a Defect_F score of 0.823, underscoring its practical application value.