Residual Learning for Salient Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Multi-scale strategies in Fully Convolutional Networks (FCNs) enhance salient object detection.
• Existing methods struggle with high-resolution mask regression and feature rescaling issues.

Purpose of the Study:

• To propose a novel residual learning strategy for salient object detection.
• To address limitations in current multi-scale deep learning approaches.

Main Methods:

• Introduced a Residual Refinement Network (R2Net) employing residual learning.
• Utilized a Dilated Convolutional Pyramid Pooling (DCPP) module for coarse predictions.
• Incorporated Attentional Residual Modules (ARMs) to guide residual learning.

Main Results:

• The R2Net progressively refines saliency maps scale-by-scale, learning residuals to correct errors.
• Achieved state-of-the-art performance on five benchmark datasets.
• Demonstrated real-time processing at 33 FPS on a single GPU without post-processing.

Conclusions:

• The proposed residual learning strategy effectively improves salient object detection.
• R2Net offers a robust, efficient, and post-processing-free solution for real-time applications.