Region-edge continuous enhancement network for salient object detection

• Keyue Yao,
• Ziguan Cui,
• Lipeng Wu,
• Zongliang Gan,
• Feng Liu,

Vol. 19, No. 3, March 31, 2025
10.3837/tiis.2025.03.008, Download Paper (Free):

• salient object detection
• Deep Learning
• Region Feature
• Edge Feature
• feature fusion

Abstract

Salient object detection (SOD) focuses on detecting the most attention-grabbing objects from an image. Despite the impressive performance of existing Fully Convolutional Network (FCN)-based methods in this domain, accurately and completely detecting salient object from complex backgrounds remains a significant issue. This paper presents a region-edge continuous enhancement network (RECE-Net) for SOD. In the proposed RECE-Net, we introduce the global-guided region enhancement (GRE) module to effectively locate salient regions and suppress non-salient background noise interference. The GRE module extracts complementary global semantic features and establishes a top-down information flow to guide each network layer in enhancing the features of salient regions. Moreover, the multi-oriented edge enhancement (OEE) module is incorporated to capture edge structure features of salient objects from multiple directions and angles, leading to improved detection accuracy by effectively learning edge information. Experiments on five benchmark SOD datasets show that our RECE-Net accurately localizes salient object regions, precisely identifies object boundaries, and outperforms previous models in terms of detection performance.

Statistics

Cite this article