EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Enhancing Visual Detection and Counting of Tiny Pests in Agricultural Environments with Dual-Granularity Textual Prompts

Zhiliang Zhang, Yu Zhang, Hongshen Guo, Yuheng Guo, Zixuan Dai, Lianyou Gui, Liang Zhang, Wei Zhan

Smart Agricultural Technology · 2026-11 · 102738

DOI: 10.1016/j.atech.2026.102738 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Smart Agricultural Technology, 2026-11. Verfasst von Zhiliang Zhang und 7 weiteren Autor:innen. Die Arbeit nennt 62 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0). Gefördert von: National Natural Science Foundation of China.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Precise field monitoring of agricultural pests is essential for scientific pest management. Existing methods mostly rely on visual information, motivating an examination of whether textual feedback can help correct an initial prediction. This study presents Text-Enhanced Pest Counter (TE-Pest), a task-focused text-conditioned framework for pest detection and counting that integrates visual information with dual-granularity textual prompts. Scene prompts provide coarse-grained semantic priors for initial detection, while optimization prompts provide fine-grained semantic guidance for result refinement. Two complementary cross-modal modules are designed. Cross-Modal Multi-Head Attention (CMHA) conditions visual features on prompt semantics through text-guided global modulation. Cross-Modal Dynamic Convolution (CMDC) generates convolution kernels dynamically from optimization prompts and visual features to adaptively refine local representations. A prompt optimization loss further guides the refinement of missed and spurious detections. Experiments were conducted on the PH-253K rice-planthopper and Thrip-47K thrips datasets. The prompted configuration is reported as a reference-assisted upper bound. It achieved F1-scores of 87.00% and 91.43% on the test set, while the autonomous prompt-free configuration achieved 85.71% and 88.93%, exceeding YOLO12-L by 7.96 percentage points on average; relative to the strongest baselines YOLO12-L and RT-DETRv2, the reference-assisted configuration reduced RMSE by approximately 34.0% and 46.4%. Within TE-Pest, dual-granularity prompts improved F1 by 1.29 and 2.50 percentage points relative to prompt-free inference. The results indicate that task-specific textual guidance provides an effective means of refining tiny-pest detection and counting predictions.

Förderung

  • National Natural Science Foundation of China

Zitieren

APA

Zhang, Z., Zhang, Y., Guo, H., Guo, Y., Dai, Z., Gui, L., Zhang, L., & Zhan, W. (2026). Enhancing Visual Detection and Counting of Tiny Pests in Agricultural Environments with Dual-Granularity Textual Prompts. Smart Agricultural Technology, 102738. https://doi.org/10.1016/j.atech.2026.102738

Vancouver

Zhang Z, Zhang Y, Guo H, Guo Y, Dai Z, Gui L, et al. Enhancing Visual Detection and Counting of Tiny Pests in Agricultural Environments with Dual-Granularity Textual Prompts. Smart Agricultural Technology. 2026:102738. doi:10.1016/j.atech.2026.102738

BibTeX
@article{Zhang2026,
  author = {Zhang, Zhiliang and Zhang, Yu and Guo, Hongshen and Guo, Yuheng and Dai, Zixuan and Gui, Lianyou and Zhang, Liang and Zhan, Wei},
  title = {Enhancing Visual Detection and Counting of Tiny Pests in Agricultural Environments with Dual-Granularity Textual Prompts},
  journal = {Smart Agricultural Technology},
  year = {2026},
  pages = {102738},
  doi = {10.1016/j.atech.2026.102738}
}
Kostenlos starten →