EDN: Salient Object Detection via Extremely-Downsampled Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient object detection (SOD) typically relies on multi-scale learning, fusing high-level and low-level features.
• Existing SOD methods predominantly focus on low-level feature enhancement, neglecting the potential of high-level features.
• High-level features, crucial for other tasks, remain under-explored in the context of SOD.

Purpose of the Study:

• To investigate the underutilized role of high-level features in salient object detection.
• To introduce a novel network architecture that effectively leverages high-level features for improved SOD performance.
• To develop a method for accurate salient object localization and fine detail recovery.

Main Methods:

• Introduced an Extremely-Downsampled Network (EDN) utilizing extreme downsampling for a global image perspective.
• Developed Scale-Correlated Pyramid Convolution (SCPC) for effective multi-level feature fusion and detail reconstruction.
• Proposed an efficient variant, EDN-Lite, for real-time SOD applications.

Main Results:

• EDN achieves state-of-the-art performance in salient object detection.
• The EDN architecture demonstrates real-time processing speeds.
• EDN-Lite achieves competitive performance with a processing speed of 316 frames per second.

Conclusions:

• Enhancing high-level features is critical for advancing salient object detection.
• The proposed EDN and EDN-Lite offer efficient and effective solutions for SOD.
• This research is expected to stimulate new directions in salient object detection methodologies.