Mixed-Supervised Scene Text Detection With Expectation-Maximization Algorithm.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Scene text detection is crucial but challenging, especially for arbitrarily-shaped text.
• Existing methods require extensive polygon-level annotations, increasing data labeling costs.
• Weakly-supervised methods often lack full object coverage.

Purpose of the Study:

• To develop a mixed-supervised learning framework for arbitrarily-shaped scene text detection.
• To reduce the reliance on costly polygon-level annotations by incorporating weak supervision.
• To approximate the performance of fully-supervised methods using reduced annotation efforts.

Main Methods:

• Proposed an Expectation-Maximization (EM) based mixed-supervised learning framework.
• Treated polygon-level labels as latent variables recovered via the EM algorithm from weak labels.
• Introduced a novel contour-based scene text detector to leverage weak labels effectively.

Main Results:

• Achieved performance comparable to fully supervised methods using only 10% strong and 90% weak annotations.
• Attained state-of-the-art results on five scene text benchmarks with 100% strong annotations.
• Demonstrated competitive performance on the ICDAR2015 dataset.

Conclusions:

• The proposed EM-based mixed-supervised framework significantly reduces annotation costs for scene text detection.
• The method offers a viable alternative to fully supervised approaches, especially when strong annotations are scarce.
• The publicly released weakly annotated datasets will benefit future research in this area.