Pixel-Level Noise Mining for Weakly Supervised Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning for visual saliency detection demands extensive, meticulously annotated datasets.
• Weakly supervised methods offer an alternative but struggle with noisy labels inherent in easily obtainable, unsupervised data.
• Deep networks are prone to overfitting noisy labels, significantly degrading performance.

Purpose of the Study:

• To develop a robust salient object detection (SOD) method using weakly supervised learning from single noisy labels.
• To address the challenge of deep network overfitting to label noise.
• To enable accurate saliency detection without relying on external models or fully supervised data.

Main Methods:

• Proposed a pixel-level noise mining framework that leverages the network's own knowledge to identify outliers.
• Introduced a progressive outlier identification process during early training stages to prevent overfitting.
• Generated an adaptive selection matrix to guide label noise rectification for improved supervision in later stages.

Main Results:

• The proposed method achieves saliency detection performance comparable to state-of-the-art fully supervised approaches.
• Outperformed existing weakly supervised methods using single noisy labels.
• Surpassed half of existing weakly supervised methods that use multiple noisy labels.
• When trained with multiple noisy labels, the method outperformed all other multi-noisy-label approaches across four datasets.
• Demonstrated generalization ability on the multiclass semantic segmentation task.

Conclusions:

• The developed noise mining framework enables robust salient object detection (SOD) from weakly supervised, noisy labels.
• The approach effectively mitigates the negative impact of label noise on deep model training.
• Achieved state-of-the-art results in weakly supervised SOD and showed strong generalization capabilities.