BiconNet: An Edge-preserved Connectivity-based Approach for Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Traditional deep learning methods for salient object detection (SOD) treat it as a pixel-wise task.
• Current SOD models often struggle with insufficient inter-pixel information, leading to imperfect segmentation at object edges and poor spatial coherence.
• Using only saliency masks as labels is suboptimal for comprehensive SOD.

Purpose of the Study:

• To address limitations in current SOD models by improving the utilization of inter-pixel information.
• To develop a novel approach that enhances segmentation accuracy, particularly near object edges and for spatial coherence.
• To introduce a flexible method that can augment existing state-of-the-art SOD frameworks.

Main Methods:

• Proposed a connectivity-based approach named bilateral connectivity network (BiconNet).
• Utilized both connectivity masks and saliency masks as labels to model inter-pixel relationships and object saliency effectively.
• Introduced a bilateral voting module for enhancing the output connectivity map and an edge feature enhancement method for utilizing edge-specific features.

Main Results:

• Demonstrated significant performance improvements across five benchmark datasets.
• Showcased the effectiveness of BiconNet in improving segmentation accuracy and spatial coherence.
• Validated that the proposed method can be integrated into existing SOD frameworks with minimal parameter increase.

Conclusions:

• BiconNet offers a superior approach to salient object detection by incorporating connectivity information.
• The method effectively models inter-pixel relationships, leading to more accurate object segmentation, especially at boundaries.
• BiconNet provides a valuable enhancement for existing SOD techniques, improving performance with negligible overhead.