Comprehensive mining of information in Weakly Supervised Semantic Segmentation: Saliency semantics and edge...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Weakly Supervised Semantic Segmentation (WSSS) using image-level labels suffers from incomplete semantic information.
• Key challenges include insufficient saliency semantic mining and neglected edge semantics.
• Existing methods struggle to comprehensively extract semantic details from limited labels.

Purpose of the Study:

• To propose a novel two-stage framework, Saliency Semantic Full Mining-Edge Semantic Mining (SSFM-ESM), for WSSS.
• To address the limitations of insufficient saliency and edge semantic information mining.
• To enhance the comprehensive information mining perspective in WSSS.

Main Methods:

• Implemented a two-stage framework: SSFM for saliency mining and ESM for edge mining.
• Utilized a pixel-level class-agnostic distance loss in the first stage (SSFM) for saliency feature learning.
• Employed an edge semantic mining module in the second stage to refine pseudo-labels and leverage network self-correction for edge information.

Main Results:

• The SSFM stage successfully mined full saliency semantic information, generating initial pseudo-labels.
• The ESM stage obtained high-confidence edge semantics by avoiding false information and utilizing network self-correction.
• Experiments on PASCAL VOC 2012 and MS COCO 2014 datasets demonstrated the approach's feasibility and superiority.

Conclusions:

• The proposed SSFM-ESM framework effectively addresses incomplete semantic information in WSSS.
• The method achieves superior performance by comprehensively mining both saliency and edge semantics.
• This work offers a promising direction for advancing WSSS techniques.