A Pixel-Based Character Detection Scheme for Texts with Arbitrary Orientations in Natural Scenes

In recent years, there has been a significant focus on deep learning-based research for detecting texts in natural scenes. While many studies have achieved promising results by targeting word detection, challenges remain in detecting and recognizing texts with arbitrary orientations. Complex image backgrounds, text occlusion, and variations in text styles easily affect the detection process of words. This paper introduces a pixel-based character detection scheme for extracting individual characters within words. The objective is to locate characters in irregular text orientations or shapes, thereby achieving better alignment of detection bounding boxes with character edges. Since existing datasets only provide word-level annotations and lack character-level ground truths, we generate realistically synthesized artificial data to address this limitation. We employ weakly supervised learning, utilizing partially annotated data for training, and subsequently enhance performance by incorporating actual data. Experimental results demonstrate that our scheme outperforms other character-level detection models regarding text recognition accuracy, as evidenced by comparisons on datasets such as ICDAR2017, TotalText, and CTW-1500.