Rethinking Lightweight Salient Object Detection via Network Depth-Width Tradeoff.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing salient object detection (SOD) methods often rely on deep, wide networks, leading to high computational costs and slow inference.
• There is a need for efficient SOD models that maintain competitive accuracy, especially for resource-constrained environments.

Purpose of the Study:

• To develop a lightweight framework for salient object detection (SOD) that achieves a favorable balance between efficiency and accuracy.
• To design novel architectural components and explore network scaling strategies for optimizing SOD performance.

Main Methods:

• Proposed a novel trilateral decoder framework that decouples the U-shape structure into three complementary branches to address semantic context dilution, spatial structure loss, and boundary detail absence.
• Introduced a Scale-Adaptive Pooling Module to achieve multi-scale receptive fields without additional learnable parameters.
• Investigated the accuracy-parameter-speed tradeoff by designing shallower and narrower models, resulting in CTD-S, CTD-M, and CTD-L versions.

Main Results:

• The proposed lightweight framework maintains competitive accuracy while significantly improving efficiency.
• The CTD-S (1.7M parameters, 125 FPS), CTD-M (12.6M parameters, 158 FPS), and CTD-L (26.5M parameters, 84 FPS) models demonstrate superior performance across five benchmarks.
• Achieved a better efficiency-accuracy balance compared to existing SOD methods.

Conclusions:

• The developed lightweight SOD framework offers a practical solution for efficient and accurate salient object detection.
• The trilateral decoder and adaptive pooling module are effective in refining segmentation details and enhancing multi-scale feature extraction.
• The study demonstrates the potential of lightweight SOD models for diverse application scenarios, from edge devices to high-performance platforms.