Deep Salient Object Detection with Contextual Information Guidance.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are vital for salient object detection.
• Existing methods inadequately integrate multi-level feature maps and side outputs.
• Effective integration requires leveraging both feature maps and side outputs across network layers.

Purpose of the Study:

• To propose a new strategy for guiding multi-level contextual information integration in CNNs.
• To enhance the accuracy of feature maps and the detail propagation of side outputs.
• To improve salient object detection performance by fully engaging feature maps and side outputs.

Main Methods:

• A novel guidance strategy for multi-level contextual information integration.
• Guiding shallower-level feature maps with deeper-level side outputs for enhanced object property learning.
• A group convolution module for generating high-discriminative feature maps, integrated into the guidance mechanism.

Main Results:

• Demonstrated effective integration of multi-level feature maps and side outputs.
• Achieved improved spatial detail propagation from deeper to shallower layers.
• Experimental validation on three benchmark datasets confirmed superior performance over state-of-the-art methods.

Conclusions:

• The proposed guidance strategy significantly enhances salient object detection in CNNs.
• The group convolution module contributes to generating more discriminative feature maps.
• The method offers a superior approach to multi-level feature integration for improved detection accuracy.