Weakly supervised salient object detection via image category annotation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning has advanced salient object detection.
• Fully supervised methods require extensive pixel-level annotations.
• Weakly supervised methods use less-costly annotations like category labels, but often suffer from inaccurate detection.

Purpose of the Study:

• To propose a novel weakly supervised salient object detection method using only category annotations.
• To address the challenge of inaccurate detection in existing category-based methods.
• To develop a strategy for generating and refining high-quality pseudo-labels.

Main Methods:

• Proposed a coarse object location network (COLN) for initial object localization using category information.
• Developed a pseudo-label generation and refinement strategy, including a quality check mechanism based on label pair consistency.
• Introduced a multi-decoder neural network (MDN) for saliency detection trained with the generated pseudo-labels.
• Implemented an iterative pseudo-label update strategy to optimize both pseudo-labels and the detection model.

Main Results:

• The proposed method effectively generates pixel-level pseudo-labels from category annotations.
• The quality check strategy successfully selects high-confidence pseudo-labels.
• The multi-decoder network and iterative optimization improved detection accuracy.
• Evaluations on four public datasets demonstrated superior performance compared to existing category-annotation-based methods.

Conclusions:

• The developed weakly supervised method offers a viable alternative to fully supervised approaches for salient object detection.
• Category annotation, when combined with effective pseudo-labeling strategies, can yield competitive detection performance.
• The proposed approach significantly advances the field of weakly supervised salient object detection.