BOTD: Bold Outline Text Detector

Recently, text detection has attracted sufficient attention in the field of computer vision and artificial intelligence. Among the existing approaches, regression-based models are limited to handle the texts with arbitrary shapes, while segmentation-based algorithms have high computational costs and suffer from the text adhesion problem. In this paper, we propose a new one-stage text detector, termed as Bold Outline Text Detector (BOTD), which is able to process the arbitrary-shaped text with low model complexity. Different from previous works, BOTD utilizes the Polar Minimum Distance (PMD) to encode the shortest distance between the center point and the contour of the text instance, and generates a Center Mask (CM) for each text instance. After learning the PMD heat map and CM map, the final results can be obtained with a simple Text Reconstruction Module (TRM). Since the CM resides within the text box exactly, the text adhesion problem is avoided naturally. Meanwhile, all the points on the text contour share the same PMD, so the complexity of BOTD is much lower than existing segmentation-based methods. Experimental results on three real-world benchmarks show the state-of-the-art performance of BOTD.