Arbitrary Shape Text Detection via Segmentation With Probability Maps.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Arbitrary shape text detection is challenging due to variations in size, aspect ratio, orientation, and shape.
• Segmentation-based methods show promise for text detection but are hindered by coarse-grained annotations and misclassified pixels.
• Existing datasets lack precise pixel-level annotations, impacting the performance of current text detection algorithms.

Purpose of the Study:

• To develop an innovative and robust segmentation-based method for accurate arbitrary shape text detection.
• To address the limitations of coarse-grained annotations in scene text detection datasets.
• To improve the performance of segmentation-based text detection by accurately classifying text pixels.

Main Methods:

• Proposed a novel segmentation-based detection method utilizing probability maps.
• Employed a Sigmoid Alpha Function (SAF) to convert boundary distances to probability maps.
• Utilized a group of probability maps and an iterative model to learn and assimilate distributions for text reconstruction, followed by region growth algorithms.

Main Results:

• Achieved state-of-the-art detection accuracy on several benchmarks.
• Demonstrated superior F-measure performance on Total-Text (88.79%), CTW1500 (85.75%), and MSRA-TD500 (88.93%) datasets with Watershed post-processing.
• Showcased promising results on multi-oriented and multilingual scene text detection datasets.

Conclusions:

• The proposed probability map-based segmentation method effectively detects arbitrary shape text instances.
• The approach overcomes challenges posed by inaccurate and coarse-grained annotations.
• The method demonstrates significant improvements in text detection accuracy and robustness across diverse datasets.