Middle-Level Feature Fusion for Lightweight RGB-D Salient Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing RGB-Depth Salient Object Detection (SOD) models often employ complex two-stream architectures.
• These models require a large number of parameters due to extensive unimodal feature extraction and multi-modal fusion, limiting practical applications.

Purpose of the Study:

• To develop a lightweight RGB-Depth SOD model with significantly reduced parameters.
• To maintain or improve detection performance despite parameter reduction.

Main Methods:

• Proposed a novel middle-level feature fusion structure for RGB-D SOD.
• Employed shallow subnetworks for initial low- and middle-level feature extraction.
• Implemented a single, specialized fusion module for middle-level features and a relation-aware module for cross-modal information capture.

Main Results:

• The proposed model achieved a lightweight design with only 3.9 million parameters.
• The model demonstrated effectiveness and superiority over state-of-the-art methods on benchmark datasets.
• Achieved a processing speed of 33 frames per second (FPS).

Conclusions:

• The novel middle-level feature fusion structure enables a lightweight yet effective RGB-D SOD model.
• The proposed method successfully balances parameter reduction with high performance in salient object detection.
• The model shows strong potential for real-world applications due to its efficiency and accuracy.