Scene Text Detection Based on Attention ConvNeXt

Abstract

The widespread applications of scene text detection have propelled it into the spotlight as a prominent area of research. However, scene text detection presents a formidable challenge in real-world scenarios, given the complexities arising from diverse backgrounds, text styles, irregular arrangements, and image blurriness. In this research, we propose a scene text detection system built upon Attention ConvNeXt. In this system, we introduce ConvNeXt V2 Tiny as the backbone architecture, with the aim of enhancing performance. Additionally, we introduce attention mechanisms, incorporate normalization methods, and modify activation functions to improve accuracy and training stability. In our experiments, our system is evaluated on three public datasets: MSRA-TD500, Total-Text, and SCUT-CTW1500. Each of these datasets is used to assess the performance of the model in different types of text regions. Experimental results demonstrate that the proposed method consistently improves the FAST baseline and achieves competitive performance with recent state-of-the-art methods while requiring substantially fewer parameters.