Large-Scale Sparse Learning From Noisy Tags for Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional semantic segmentation relies on pixel-level labels, which are labor-intensive and costly to obtain.
• Existing weakly supervised learning (WSL) methods struggle with noisy image tags, limiting segmentation accuracy.
• There is a need for robust methods that can perform semantic segmentation using weaker, less precise supervision.

Purpose of the Study:

• To develop a novel large-scale sparse learning (LSSL) approach for semantic segmentation of images with noisy tags.
• To address the limitations of traditional supervised methods by utilizing weaker image-level tag supervision.
• To formulate semantic segmentation as a noise reduction problem within a weakly supervised learning framework.

Main Methods:

• The proposed method formulates semantic segmentation as a weakly supervised learning (WSL) problem, focusing on noise reduction of superpixel labels.
• It transforms the WSL problem into a large-scale sparse learning (LSSL) problem by learning data manifolds.
• A linear-time-complexity algorithm based on nonlinear approximation and dimension reduction is developed for efficient LSSL problem solving.
• The LSSL approach is extended for visual feature refinement to further enhance semantic segmentation performance.

Main Results:

• The developed LSSL approach demonstrates effectiveness in semantic segmentation tasks using images with noisy tags.
• Experimental results show that the proposed method achieves promising segmentation accuracy despite the weak and noisy supervision.
• The noise reduction strategy within the WSL framework proves beneficial for improving segmentation outcomes.

Conclusions:

• The large-scale sparse learning (LSSL) approach offers a viable solution for semantic segmentation with noisy image tags.
• The method successfully leverages weak supervision by transforming the problem into a noise reduction task.
• The LSSL framework provides a robust and efficient way to improve semantic segmentation performance in challenging real-world scenarios.