Polygon-Aware Detection of Arbitrary-Shaped Text in Natural Scenes

Abstract

Scene text detection in natural images is a topic of significant discourse as it plays a vital role in several applications. Its uses can range from autonomous navigation to assistive technologies to augmented reality. Traditionally, pipelines for text detection relied on rectangular or rotated bounding boxes to represent text, but this method has shown limitations in capturing irregular shapes found in real-world environments. This paper represents a framework designed to address this challenge. The framework uses polygonal boxes instead of rectangular ones. This polygon-aware approach leads to more accurate localization of complex text shapes and structures. This system combines polygon boundary estimation with deep convolutional feature extraction to identify text instances across diverse shapes and orientations. The framework is evaluated on three widely adopted curved-text benchmarks - Total-Text, CTW1500, and ICDAR. The polygon-based representation yields measurable gains in detecting curved and irregular text. The approach delivers better accuracy without sacrificing computational efficiency, making it viable for real-world scene text detection.