Semisupervised Semantic Segmentation by Improving Prediction Confidence.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• High-quality pixel-level annotations are crucial for state-of-the-art image segmentation but are costly and time-consuming to acquire.
• Existing semisupervised and weakly supervised methods often struggle with prediction confidence, particularly in complex regions.

Purpose of the Study:

• To develop a semisupervised semantic segmentation approach that minimizes the need for extensive pixel-level annotations.
• To improve the confidence and accuracy of segmentation predictions, especially in challenging boundary areas.

Main Methods:

• An adversarial framework was employed, treating the segmentation network as a generator and a fully convolutional network as a discriminator.
• Information entropy was utilized to quantify prediction uncertainty, enabling targeted refinement of misclassified regions in unlabeled data.

Main Results:

• The proposed method demonstrated enhanced confidence in predicted class probability maps.
• Experimental results on PASCAL VOC 2012 and PASCAL-CONTEXT datasets showed competitive segmentation performance.

Conclusions:

• The novel semisupervised approach effectively improves segmentation prediction confidence by focusing on uncertain and misclassified regions.
• This method offers a viable alternative to fully supervised learning, reducing annotation costs while maintaining competitive performance.