Starnet-tex: lightweight real-time text detection via middle-order enhanced contextual attention
Hu Zheng, Yingnan Zhao, Zuguo Yang, Jielin Jiang, Xiaolong Xu · Complex & Intelligent Systems · 2025
In recent years, scene text detection has garnered significant academic attention due to its broad applications. Existing methods still struggle to efficiently achieve precise localization of arbitrary-shaped text in complex natural scenes, particularly facing notable challenges in balancing detection accuracy and inference speed. To address this issue, this paper proposes StarNet-Tex, a lightweight real-time scene text detection framework. First, we propose a Contextual Attention Tail module (CAT) to address the long-range dependency deficiencies in shallow backbone networks. It achieves this through synergistic multi-directional large-kernel receptive field expansion and local feature enhancement, ultimately constructing the lightweight backbone network StarCAN. Second, we design a Channel Aggregation Feature Pyramid (CAFP) with cascaded U-shaped structures, which enhances middle-order feature representation through dynamic balancing of low- and high-order interactions. Furthermore, we introduce a Spatial-Channel Reconstruction Module (SCRM) to suppress background interference in dense text scenarios. This module employs cross-reconstruction strategies and channel-spatial decoupling mechanisms, thereby effectively reducing false detection rates. Experimental results demonstrate that StarNet-Tex achieves F-measure scores of 85.1%, 86.5%, and 85.8% on the CTW1500, Total-Text, and ICDAR2015 datasets, respectively. With only 3.3M parameters and real-time inference at 50.9 FPS, our framework provides a novel solution that synergistically optimizes accuracy and efficiency for text detection in resource-constrained scenarios. The code will be available at https://github.com/radish512/StarNet-Tex .