EUVIMEDEUVIMED entdecken
Crossref

Multi-query text spotter transformer for arbitrary shape scene text detection and recognition

Canhui Xu, Yaqi Chen, Jialiang Li, Cao Shi

Physica Scripta · 14.08.2026 · 101 (32) · 325002

DOI: 10.1088/1402-4896/ae94bb ↗

Kurz gefasst

Fachartikel in der Zeitschrift Physica Scripta, online erschienen am 12.08.2026. Verfasst von Canhui Xu und 3 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Control points regression based detection task, which predicts Bezier curve or polygon control points to locate text position, was considered preliminary step for scene text recognition. Recent Transformer based methods jointly connect the detection and recognition within one unified network. To explore the task relational contexts and describe the mutual effects, in this work, we propose a novel Multi-Query TExt Spotter TRansformer (Multi-Query TESTR), which consists of a shared transformer encoder and a single Multi-Query Transformer Decoder. Both detection and recognition tasks shared the same transformer encoder to process the multi-scale inputs with positional embedding. Detection queries focus on spatial awareness across multiple feature scales to regress control points. Conversely, recognition queries capture the semantic features of text instances. To learn the mutual effects between these task related queries, the cross task attention is obtained by self attention transformer for reasoning dependencies among these tasks. With the fused interaction of multiple queries, two separate branches are designed to accomplish the specific detection and recognition query learning. After concatenation of the two branches, Collaborative Multi-Scale Deformable Cross Attention is performed to learn the cross attention between queries and feature contents. Experiments conducted on three publicly available benchmarks SCUT-CTW1500, Total-Text, and ICDAR15 have demonstrated that Multi Query TESTR achieves state-of-the-art performance on arbitrary shaped and multi-oriented scene text detection and recognition.

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Xu, C., Chen, Y., Li, J., & Shi, C. (2026). Multi-query text spotter transformer for arbitrary shape scene text detection and recognition. Physica Scripta, 101(32), 325002. https://doi.org/10.1088/1402-4896/ae94bb

Vancouver

Xu C, Chen Y, Li J, Shi C. Multi-query text spotter transformer for arbitrary shape scene text detection and recognition. Physica Scripta. 2026;101(32):325002. doi:10.1088/1402-4896/ae94bb

BibTeX
@article{Xu2026,
  author = {Xu, Canhui and Chen, Yaqi and Li, Jialiang and Shi, Cao},
  title = {Multi-query text spotter transformer for arbitrary shape scene text detection and recognition},
  journal = {Physica Scripta},
  year = {2026},
  volume = {101},
  number = {32},
  pages = {325002},
  doi = {10.1088/1402-4896/ae94bb}
}
Kostenlos starten →