Semi-Supervised Pixel-Level Scene Text Segmentation by Mutually Guided Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Scene text segmentation is crucial for image understanding but lacks sufficient pixel-level labeled data for supervised learning.
• Existing methods for scene text detection (region masks) are more established than pixel-level segmentation.

Purpose of the Study:

• To develop a data-driven method for accurate pixel-level scene text segmentation from single natural images.
• To address the challenge of limited pixel-level labeled data by leveraging readily available text region masks.

Main Methods:

• Propose a mutually guided network with two branches: one for polygon-level masks and another for pixel-level text masks.
• Employ a semi-supervised learning strategy where the outputs of both branches guide each other during training.
• Utilize text region masks as auxiliary data to compensate for the scarcity of pixel-level annotations.

Main Results:

• The proposed mutually guided network achieves state-of-the-art performance in pixel-level scene text segmentation.
• Experimental results validate the effectiveness of the network in accurately segmenting text at the pixel level.
• The generated segmentation masks demonstrate potential for improving downstream text recognition tasks.

Conclusions:

• The novel mutually guided network effectively performs pixel-level scene text segmentation, overcoming data scarcity challenges.
• Leveraging text region masks as auxiliary data is a viable strategy for improving segmentation accuracy.
• The method shows promise beyond segmentation, enhancing text recognition capabilities.