Deeply Supervised Salient Object Detection with Short Connections

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) have driven progress in salient object detection.
• Current methods based on Fully Convolutional Networks (FCNs) often neglect the scale-space problem.
• Existing architectures like Holistically-Nested Edge Detector (HED) show limited gains in saliency detection.

Purpose of the Study:

• To propose a novel salient object detection method.
• To address limitations of generic FCN models in handling scale-space variations.
• To improve the effectiveness and efficiency of salient object detection.

Main Methods:

• Introduced short connections to the skip-layer structures within the HED architecture.
• Leveraged multi-level and multi-scale features from FCNs for advanced representations.
• Developed a framework integrating enhanced skip-layer structures for segment detection.

Main Results:

• Achieved state-of-the-art performance on five benchmark datasets.
• Demonstrated advantages in efficiency (0.08 seconds per image), effectiveness, and simplicity.
• Conducted an analysis on the impact of training data and provided a new training set.

Conclusions:

• The proposed method significantly enhances salient object detection capabilities.
• The integration of short connections improves feature representation and detection accuracy.
• The framework offers a simple yet powerful solution for salient object detection, suitable for future research.