Salient Object Detection via Integrity Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current salient object detection (SOD) models struggle with the integrity of predicted salient regions.
• Integrity is defined at micro (part completeness) and macro (all objects detected) levels.
• Existing SOD methods often prioritize feature discriminability over integral object representation.

Purpose of the Study:

• To introduce a novel Integrity Cognition Network (ICON) for enhancing salient object detection integrity.
• To address limitations in micro-level (part completeness) and macro-level (all objects detected) salient region prediction.
• To improve the discovery and accurate delineation of all salient objects within an image.

Main Methods:

• Developed the Integrity Cognition Network (ICON) with three key components: Diverse Feature Aggregation (DFA), Integrity Channel Enhancement (ICE), and Part-Whole Verification (PWV).
• DFA aggregates features with diverse receptive fields to increase feature diversity for integral object mining.
• ICE enhances salient object-related feature channels while suppressing distractors, and PWV verifies part-whole feature agreement for micro-level integrity.

Main Results:

• ICON demonstrated superior performance across seven challenging benchmarks compared to baseline SOD methods.
• Achieved approximately a 10% relative improvement in average false negative ratio (FNR) on six datasets compared to the previous best model.
• The proposed components effectively enhance feature diversity, channel relevance, and part-whole consistency for improved SOD.

Conclusions:

• The Integrity Cognition Network (ICON) effectively addresses the integrity limitations in salient object detection.
• ICON's novel components significantly improve both micro and macro level integrity for salient object prediction.
• The model shows state-of-the-art performance, offering substantial gains in accuracy and reducing detection errors.