Hierarchical Edge Refinement Network for Saliency Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Current saliency detection methods, primarily based on fully convolutional neural networks (FCNs), often produce blurred object edges due to limited spatial resolution from convolution and pooling operations.
• Accurate edge delineation is crucial for precise salient object detection, a common task in computer vision.

Purpose of the Study:

• To address the edge blurring issue in FCN-based saliency detection.
• To propose a novel Hierarchical Edge Refinement Network (HERNet) for accurate salient object detection with refined edges.

Main Methods:

• Developed a Hierarchical Edge Refinement Network (HERNet) comprising a saliency prediction network (modified U-Net) and an edge preserving network (utilizing the atrous spatial pyramid pooling (ASPP) module).
• Implemented a novel one-to-one hierarchical supervision strategy to effectively train the network.
• Evaluated performance on five traditional benchmark datasets.

Main Results:

• The proposed HERNet successfully overcomes the edge blurring problem inherent in traditional FCN methods.
• Experimental results demonstrate that HERNet achieves competitive performance compared to state-of-the-art saliency detection methods.
• The network effectively predicts salient object regions and preserves fine edge details.

Conclusions:

• HERNet offers a significant advancement in accurate salient object detection by focusing on edge refinement.
• The hierarchical structure and supervision strategy contribute to improved performance and edge accuracy.
• The proposed method shows strong potential for applications requiring precise object boundary detection.