IMSFNet: integrated multi-source feature network for salient object detection

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 53; no. 19; pp. 22228 - 22248
• Main Authors: Xia, Chenxing, Sun, Yanguang, Fang, Xianjin, Ge, Bin, Gao, Xiuju, Li, Kuan-Ching
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.10.2023; Springer Nature B.V
• Subjects: Artificial Intelligence; Computer Science; Context; Deep learning; Machines; Manufacturing; Mechanical Engineering; Methods; Modules; Object recognition; Performance enhancement; Processes; Salience; Semantics; Dilated convolution; Context information; Salient object detection; Multi-scale feature
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-023-04636-y

Multi-scale context features are conducive to image understanding, so it plays an important role in salient object detection (SOD) tasks, and contextual information-based SOD methods have achieved fine performance. However, the context information obtained through parallel independent convolutions with large kernels or dilated convolutions with different dilation rates is lack relevance and dependence at different scales, which limits the expressive ability of context information. In this article, we propose a novel Integrated Multi-Source Feature Network (IMSFNet) for accurate SOD task, which mainly consists of three components. Specifically, we first develop a multi-scale feature aggregation module (MSFAM) to adequately capture and utilize multi-scale context features through a series of well-designed dilated convolutions and short hierarchical connections, and then aggregate these information to improve the performance of input initial features. Subsequently, based on the extracted high-level features, we introduce a global feature extractor (GFE) to further excavate higher-level global semantic information to help locate salient objects from cluttered real-world scenes. Finally, a correlation feature interaction module (CFIM) is designed to interact the diverse information from different level features, reducing the interference of complex backgrounds and highlighting salient objects. Extensive experimental results on six public SOD benchmark datasets convincingly demonstrate the effectiveness and superiority of the proposed IMSFNet method against the 18 state-of-the-art SOD methods under different evaluation metrics.