Large-Scale Insect Detection With Fine-Tuning YOLOX

Abstract

With the aim of detecting insect pests at an early stage, there has been an increasing demand for insect pest detection and classification, particularly in large-scale setups. Therefore, the aim of this research is to introduce a new real-time pest detection technique using a deep convolutional neural network, which not only offers improved accuracy but also faster speed and less computational effort. The networks were constructed using various modern object detector models such as YOLOv4, YOLOv5, and YOLOX. Our proposed networks were evaluated on a standard large-scale insect pest dataset, IP102, as well as on our collected dataset, Insect10. The experimental results demonstrate that our system surpasses previous methods and achieves satisfactory performance with 84.84% mAP on the Insect10 dataset and 54.19% mAP on the IP102 dataset. Our system can deliver precise and real-time pest detection and identification for agricultural crops, enabling highly accurate end-to-end pest detection that can be applied in realistic farming scenarios.