Proposal of an advanced YOLOX model for real-time detection of Vespa hornets (Hymenoptera; Vespidae), key pests of honey bees

• Published in: Journal of Asia-Pacific entomology Vol. 27; no. 2; pp. 102234 - 8
• Main Authors: Kwon, Youngjae, Lee, Cheolhee, Bak, Seongbin, Jung, Chuleui
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2024; 한국응용곤충학회
• Subjects: data collection; Deep-learning; ecosystems; entomology; forests; honey; honey bees; Object detection; pest control; Real-time vespa detection; social wasps; Vespa; YOLOX; 농학; YOLOX; Deep-learning; Object detection; Real-time vespa detection
• ISSN: 1226-8615; 1876-7990
• DOI: 10.1016/j.aspen.2024.102234

[Display omitted] •Deep learning based on real-time monitoring technology is developed.•An advanced YOLOX model for real-time detection of Vespa hornets was proposed.•New model enhanced the discrimination power.•Process speed also was shorter than the conventional model.•This model has application in Vespa monitoring, pest control, and insect detection. Deep learning based on real-time monitoring of important honeybee pests would contribute to enhancing honeybee health and ecosystem services honey bee provides. This study introduced an advanced YOLOX model for Vespa hornet monitoring and evaluated its detection performance compared to vanilla YOLOX. A new feature extraction layer based on the shuffle layer is introduced to enhance detection performance for small objects such as Vespa while reducing inference speed simultaneously. Five sets of experiments were conducted. Detection precision (mAP@50) showed 93.8% for the proposed model and 89.6% for the vanilla YOLOX model. Second, a data generalization test showed comparable performance to the dataset of 31 forest pests. Third, a detection accuracy test for small objects is done. For this test, we preprocessed all images of the custom Vespa dataset to obtain resized Vespa objects with a 0.3% size ratio for all 900 test images, in which the average pixel size ratio between each Vespa object’s pixel size to the corresponding whole image’s pixel size is set to approximately 0.3%. The proposed structure also showed better results by 1.14% in terms of mAP@50 in experiment 3 for 0.3% test Vespa images. Fourth and fifth experiments were inference time comparisons for 0.3% test Vespa image dataset of experiment 3 and FHD-sized webcam images. In terms of inference speed, the proposed model was 1.34–1.35 times faster than the vanilla YOLOX model due to optimized convolution operation in the backbone. Therefore, our advanced model for YOLOX is verified to be more effective than the state-of-the-art YOLOX model in terms of inference accuracy and speed. It can be widely applied in fields of small object detection, such as Vespa monitoring, pest control, and insect detection.