ETLRS: An End-to-End Text Spotting Framework With Spatial and Directional Perception and Cross-Task Feature Synergy

Text spotting in natural scenes is crucial to numerous applications, including document understanding, autonomous driving, navigation systems, and intelligent content auditing. Although recent advances in Optical Character Recognition (OCR) and scene text detection have significantly improved performance in controlled environments, end-to-end text spotting in complex real-world scenes remains challenging. Issues such as irregular text shapes, diverse orientations, cluttered backgrounds, and varying illumination conditions often lead to unstable detection and inaccurate recognition. Furthermore, conventional stepwise pipelines fail to leverage the intrinsic correlations between detection and recognition, limiting the robustness and generalization of the overall system. To address these challenges, this paper proposes ETLRS, an end-to-end text spotting framework that enhances text perception and strengthens feature-level collaboration between detection and recognition. First, a Spatial–Directional Perception (SDP) module is introduced to enrich text region features by integrating spatial structure cues and directional edge information, thereby improving robustness against shape deformation and visual noise. Second, a Cross-Task Feature Synergy (CTFS) module is designed to explicitly align geometric cues from localization with semantic cues from recognition, generating synergistic representations that jointly benefit both detection and recognition. Finally, a query-guided decoder incorporates these synergistic features to simultaneously produce text localization and transcription outputs in an end-to-end manner. Extensive experiments on Total-Text, ICDAR2015, and CTW1500 demonstrate that ETLRS consistently improves text detection and end-to-end spotting, achieving the best results on Total-Text and ICDAR2015 and strong robustness on CTW1500. ETLRS also outperforms the recent state-of-the-art FastTextSpotter on Total-Text and ICDAR2015 and delivers clear gains under lexicon-free evaluation on CTW1500. Meanwhile, ETLRS is lightweight with 40.9M parameters and runs at 21.3 FPS, achieving the fastest inference among the compared methods.