AUTHOR=Liu Jingwei , Li Yongke , Wang Lei , Zhao Yunjie , Mao Bowen , Wang Pengying 

TITLE=GIWT-YOLO: an efficient multi-scale framework for real-time Scolytinae pests detection

JOURNAL=Frontiers in Insect Science

VOLUME=Volume 5 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/insect-science/articles/10.3389/finsc.2025.1635439

DOI=10.3389/finsc.2025.1635439

ISSN=2673-8600

ABSTRACT=The broad range of Scolytinae pests sizes and their subtle visual similarities, especially in smaller species, continue to challenge the accuracy of mainstream object detection models. To address these challenges, this paper proposes GIWT-YOLO, a lightweight detection model based on YOLOv11s, specifically tailored for Scolytinae pests detection. (1) We designed a lightweight multi-scale convolution module, GIConv, to improve the model’s ability to extract features at different pest scales. This module enhances the accuracy of small-object detection while reducing the computational cost and parameter complexity of the backbone. (2) The WTConv module inspired by wavelet transform is introduced into the backbone. This enlarges the effective receptive field and improves the model’s ability to distinguish pests with similar textures. (3) An SE attention mechanism is incorporated between the Neck and Head to enhance the model’s focus on key feature regions. Experimental results show that GIWT-YOLO achieves 84.7% in Precision, 88.7% in mAP@50, and 63.4% in mAP@50~95, which are improvements of 2.2%, 4.0%, and 3.1%, respectively, compared to the baseline YOLOv11s. Additionally, the model’s parameters and GFLOPs are reduced by 11.3% and 13.4%, respectively. Our proposed model surpasses the state-of-the-art (SOTA) performance in small-sized pest detection while maintaining a lightweight architecture, and its generalization ability has been validated on other public datasets. Our model provides an efficient solution for detecting Scolytinae pests. In future work, we plan to collect additional images of various pest species to expand the dataset, further enhancing the model’s applicability to a wider range of pest detection scenarios.