Progressive Self-Guided Loss for Salient Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Current salient object detection (SOD) models struggle with incomplete predictions due to object complexity.
• Existing methods often fail to fully capture the intricate details of salient objects.

Purpose of the Study:

• To develop a novel progressive self-guided loss function for enhancing deep learning-based SOD.
• To improve the completeness and spatial accuracy of saliency maps.
• To introduce a new feature aggregation module for adaptive multi-scale feature utilization.

Main Methods:

• A progressive self-guided loss function simulating morphological closing operations is proposed.
• Auxiliary training supervisions are progressively generated to guide the model.
• A feature aggregation module with a branch-wise attention mechanism captures and integrates multi-scale features.
• The approach is evaluated on benchmark datasets for salient object detection.

Main Results:

• The proposed loss function guides SOD models to produce more complete salient object predictions.
• It effectively uncovers spatial dependencies among salient object pixels using a region growing approach.
• The feature aggregation module enhances the framework's ability to locate and detect salient objects.
• State-of-the-art performance is achieved on benchmark datasets.

Conclusions:

• The progressive self-guided loss function significantly improves salient object detection without altering existing model architectures.
• The novel feature aggregation module contributes to effective multi-scale feature utilization.
• The proposed methods advance the state-of-the-art in deep learning-based salient object detection.